llama-chat : reset glmedge chat template (#13253)

* reset glmedge chat template

* fix glmedge chat template

.clang-tidy

@@ -13,6 +13,7 @@ Checks: >
     -readability-magic-numbers,
     -readability-uppercase-literal-suffix,
     -readability-simplify-boolean-expr,
+    -readability-math-missing-parentheses,
     clang-analyzer-*,
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,

.devops/cpu.Dockerfile

@@ -14,9 +14,9 @@ WORKDIR /app
 COPY . .
 
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
     elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
+        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \

.devops/cuda.Dockerfile

@@ -21,7 +21,7 @@ COPY . .
 RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
     export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/intel.Dockerfile

@@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
     echo "Building with dynamic libs" && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DLLAMA_BUILD_TESTS=OFF ${OPT_SYCL_F16} && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/llama-cli-cann.Dockerfile

@@ -22,7 +22,7 @@ ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
 
 RUN echo "Building with static libs" && \
     source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF -DLLAMA_BUILD_TESTS=OFF  && \
     cmake --build build --config Release --target llama-cli
 
 # TODO: use image with NNRT

.devops/musa.Dockerfile

@@ -35,7 +35,7 @@ COPY . .
 RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \
         export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \
     fi && \
-    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.devops/rocm.Dockerfile

@@ -40,7 +40,7 @@ WORKDIR /app
 COPY . .
 
 RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \
+    cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_BUILD_TESTS=OFF \
     && cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib \

.devops/vulkan.Dockerfile

@@ -16,7 +16,7 @@ WORKDIR /app
 
 COPY . .
 
-RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
+RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1  -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \
     cmake --build build --config Release -j$(nproc)
 
 RUN mkdir -p /app/lib && \

.github/workflows/build-linux-cross.yml

@@ -4,18 +4,25 @@ on:
   workflow_call:
 
 jobs:
-  ubuntu-latest-riscv64-cpu-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-cpu-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
       - name: Setup Riscv
         run: |
           sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
@@ -40,21 +47,25 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-latest-riscv64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-riscv64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
       - name: Setup Riscv
         run: |
           sudo dpkg --add-architecture riscv64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \
@@ -82,21 +93,25 @@ jobs:
 
           cmake --build build --config Release -j $(nproc)
 
-  ubuntu-latest-arm64-vulkan-cross:
-    runs-on: ubuntu-latest
+  ubuntu-24-arm64-vulkan-cross:
+    runs-on: ubuntu-24.04
 
     steps:
       - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-
       - name: Setup Arm64
         run: |
           sudo dpkg --add-architecture arm64
-          sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \
-                 /etc/apt/sources.list /etc/apt/apt-mirrors.txt
-          sudo apt-get clean
-          sudo apt-get update
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/arm64-ports.list
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \

.github/workflows/build.yml

@@ -601,9 +601,8 @@ jobs:
             -DGGML_SYCL_F16=ON
           cmake --build build --config Release -j $(nproc)
 
-# Disabled for now due to sporadic issue syncing.
-#  build-linux-cross:
-#    uses: ./.github/workflows/build-linux-cross.yml
+  build-linux-cross:
+    uses: ./.github/workflows/build-linux-cross.yml
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest

README.md

@@ -16,9 +16,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
-- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
-- **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
+- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
+- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
+- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
 - Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123

SECURITY.md

@@ -40,7 +40,8 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks
 
 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value
+* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
+* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
 * Encrypt your data if sending it over the network.
 
 ### Multi-Tenant environments

cmake/build-info.cmake

@@ -41,14 +41,20 @@ endif()
 
 if(MSVC)
     set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    if (CMAKE_VS_PLATFORM_NAME)
+        set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
+    else()
+        set(BUILD_TARGET "${CMAKE_SYSTEM_NAME} ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
 else()
     execute_process(
-        COMMAND sh -c "\"$@\" --version | head -1" _ ${CMAKE_C_COMPILER}
+        COMMAND ${CMAKE_C_COMPILER} --version
         OUTPUT_VARIABLE OUT
         OUTPUT_STRIP_TRAILING_WHITESPACE
     )
+    string(REGEX REPLACE " *\n.*" "" OUT "${OUT}")
     set(BUILD_COMPILER ${OUT})
+
     execute_process(
         COMMAND ${CMAKE_C_COMPILER} -dumpmachine
         OUTPUT_VARIABLE OUT

common/CMakeLists.txt

@@ -39,7 +39,9 @@ add_custom_command(
     COMMENT "Generating build details from Git"
     COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
             -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
+            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
+            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
     VERBATIM

common/arg.cpp

@@ -38,6 +38,30 @@
 
 using json = nlohmann::ordered_json;
 
+std::initializer_list<enum llama_example> mmproj_examples = {
+    LLAMA_EXAMPLE_LLAVA,
+    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
+};
+
+static std::string read_file(const std::string & fname) {
+    std::ifstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    file.close();
+    return content;
+}
+
+static void write_file(const std::string & fname, const std::string & content) {
+    std::ofstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    file << content;
+    file.close();
+}
+
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
     this->examples = std::move(examples);
     return *this;
@@ -157,6 +181,10 @@ struct common_hf_file_res {
 
 #ifdef LLAMA_USE_CURL
 
+bool common_has_curl() {
+    return true;
+}
+
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
@@ -189,11 +217,11 @@ struct curl_slist_ptr {
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
     int remaining_attempts = max_attempts;
 
     while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
 
         CURLcode res = curl_easy_perform(curl);
         if (res == CURLE_OK) {
@@ -204,6 +232,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
         LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
 
         remaining_attempts--;
+        if (remaining_attempts == 0) break;
         std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
     }
 
@@ -222,8 +251,6 @@ static bool common_download_file_single(const std::string & url, const std::stri
         return false;
     }
 
-    bool force_download = false;
-
     // Set the URL, allow to follow http redirection
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
@@ -247,7 +274,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
+    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
     std::string etag;
     std::string last_modified;
 
@@ -257,14 +284,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         if (metadata_in.good()) {
             try {
                 metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
+                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
                 if (metadata.contains("etag") && metadata.at("etag").is_string()) {
                     etag = metadata.at("etag");
                 }
@@ -272,10 +292,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
                     last_modified = metadata.at("lastModified");
                 }
             } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
             }
         }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
     } else {
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
@@ -287,7 +307,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
     };
 
     common_load_model_from_url_headers headers;
+    bool head_request_ok = false;
+    bool should_download = !file_exists; // by default, we should download if the file does not exist
 
+    // get ETag to see if the remote file has changed
     {
         typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
@@ -316,23 +339,28 @@ static bool common_download_file_single(const std::string & url, const std::stri
         curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        // we only allow retrying once for HEAD requests
+        // this is for the use case of using running offline (no internet), retrying can be annoying
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
         if (!was_perform_successful) {
-            return false;
+            head_request_ok = false;
         }
 
         long http_code = 0;
         curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        if (http_code == 200) {
+            head_request_ok = true;
+        } else {
+            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+            head_request_ok = false;
         }
     }
 
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
+    // if head_request_ok is false, we don't have the etag or last-modified headers
+    // we leave should_download as-is, which is true if the file does not exist
+    if (head_request_ok) {
+        // check if ETag or Last-Modified headers are different
+        // if it is, we need to download the file again
         if (!etag.empty() && etag != headers.etag) {
             LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
             should_download = true;
@@ -341,6 +369,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
             should_download = true;
         }
     }
+
     if (should_download) {
         std::string path_temporary = path + ".downloadInProgress";
         if (file_exists) {
@@ -394,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         // start the download
         LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
             llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
         if (!was_perform_successful) {
             return false;
         }
@@ -415,13 +444,15 @@ static bool common_download_file_single(const std::string & url, const std::stri
             {"etag", headers.etag},
             {"lastModified", headers.last_modified}
         });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+        write_file(metadata_path, metadata.dump(4));
+        LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
 
         if (rename(path_temporary.c_str(), path.c_str()) != 0) {
             LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
             return false;
         }
+    } else {
+        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
     }
 
     return true;
@@ -522,6 +553,50 @@ static bool common_download_model(
     return true;
 }
 
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params) {
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::vector<char> res_buffer;
+
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        auto data_vec = static_cast<std::vector<char> *>(data);
+        data_vec->insert(data_vec->end(), (char *)ptr, (char *)ptr + size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_buffer);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (params.timeout > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_TIMEOUT, params.timeout);
+    }
+    if (params.max_size > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size);
+    }
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    for (const auto & header : params.headers) {
+        http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str());
+    }
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        std::string error_msg = curl_easy_strerror(res);
+        throw std::runtime_error("error: cannot make GET request: " + error_msg);
+    }
+
+    long res_code;
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+
+    return { res_code, std::move(res_buffer) };
+}
+
 /**
  * Allow getting the HF file from the HF repo with tag (like ollama), for example:
  * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
@@ -541,46 +616,48 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
         throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
     }
 
-    // fetch model info from Hugging Face Hub API
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
+    std::string url = get_model_endpoint() + "v2/" + hf_repo + "/manifests/" + tag;
 
-    std::string model_endpoint = get_model_endpoint();
-
-    std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
+    // headers
+    std::vector<std::string> headers;
+    headers.push_back("Accept: application/json");
     if (!bearer_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + bearer_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+        headers.push_back("Authorization: Bearer " + bearer_token);
     }
     // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+    // User-Agent header is already set in common_remote_get_content, no need to set it here
 
-    CURLcode res = curl_easy_perform(curl.get());
+    // we use "=" to avoid clashing with other component, while still being allowed on windows
+    std::string cached_response_fname = "manifest=" + hf_repo + "=" + tag + ".json";
+    string_replace_all(cached_response_fname, "/", "_");
+    std::string cached_response_path = fs_get_cache_file(cached_response_fname);
 
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
+    // make the request
+    common_remote_params params;
+    params.headers = headers;
+    long res_code = 0;
+    std::string res_str;
+    bool use_cache = false;
+    try {
+        auto res = common_remote_get_content(url, params);
+        res_code = res.first;
+        res_str = std::string(res.second.data(), res.second.size());
+    } catch (const std::exception & e) {
+        LOG_WRN("error: failed to get manifest: %s\n", e.what());
+        LOG_WRN("try reading from cache\n");
+        // try to read from cache
+        try {
+            res_str = read_file(cached_response_path);
+            res_code = 200;
+            use_cache = true;
+        } catch (const std::exception & e) {
+            throw std::runtime_error("error: failed to get manifest (check your internet connection)");
+        }
     }
+    std::string ggufFile;
+    std::string mmprojFile;
 
-    long res_code;
-    std::string ggufFile   = "";
-    std::string mmprojFile = "";
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
-    if (res_code == 200) {
+    if (res_code == 200 || res_code == 304) {
         // extract ggufFile.rfilename in json, using regex
         {
             std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
@@ -597,6 +674,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
                 mmprojFile = match[1].str();
             }
         }
+        if (!use_cache) {
+            // if not using cached response, update the cache file
+            write_file(cached_response_path, res_str);
+        }
     } else if (res_code == 401) {
         throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
     } else {
@@ -613,6 +694,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
 
 #else
 
+bool common_has_curl() {
+    return false;
+}
+
 static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
     LOG_ERR("error: built without CURL, cannot download model from internet\n");
     return false;
@@ -635,17 +720,30 @@ static struct common_hf_file_res common_get_hf_file(const std::string &, const s
     return {};
 }
 
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params &) {
+    if (!url.empty()) {
+        throw std::runtime_error("error: built without CURL, cannot download model from the internet");
+    }
+
+    return {};
+}
+
 #endif // LLAMA_USE_CURL
 
 //
 // utils
 //
 
-static void common_params_handle_model(
+struct handle_model_result {
+    bool found_mmproj = false;
+    common_params_model mmproj;
+};
+
+static handle_model_result common_params_handle_model(
         struct common_params_model & model,
         const std::string & bearer_token,
-        const std::string & model_path_default,
-        bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
+        const std::string & model_path_default) {
+    handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
         if (!model.hf_repo.empty()) {
@@ -657,7 +755,12 @@ static void common_params_handle_model(
                         exit(1); // built without CURL, error message already printed
                     }
                     model.hf_repo = auto_detected.repo;
-                    model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
+                    model.hf_file = auto_detected.ggufFile;
+                    if (!auto_detected.mmprojFile.empty()) {
+                        result.found_mmproj   = true;
+                        result.mmproj.hf_repo = model.hf_repo;
+                        result.mmproj.hf_file = auto_detected.mmprojFile;
+                    }
                 } else {
                     model.hf_file = model.path;
                 }
@@ -694,6 +797,8 @@ static void common_params_handle_model(
             exit(1);
         }
     }
+
+    return result;
 }
 
 const std::vector<ggml_type> kv_cache_types = {
@@ -827,16 +932,25 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    common_params_handle_model(params.model,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model(params.speculative.model, params.hf_token, "");
-    common_params_handle_model(params.vocoder.model,     params.hf_token, "");
-
-    // allow --mmproj to be set from -hf
-    // assuming that mmproj is always in the same repo as text model
-    if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
-        params.mmproj.hf_repo = params.model.hf_repo;
+    // handle model and download
+    {
+        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH);
+        if (params.no_mmproj) {
+            params.mmproj = {};
+        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
+            // optionally, handle mmproj model when -hf is specified
+            params.mmproj = res.mmproj;
+        }
+        // only download mmproj if the current example is using it
+        for (auto & ex : mmproj_examples) {
+            if (ctx_arg.ex == ex) {
+                common_params_handle_model(params.mmproj,    params.hf_token, "");
+                break;
+            }
+        }
+        common_params_handle_model(params.speculative.model, params.hf_token, "");
+        common_params_handle_model(params.vocoder.model,     params.hf_token, "");
     }
-    common_params_handle_model(params.mmproj,            params.hf_token, "", true);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -968,7 +1082,6 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
         "llama-embedding",
         "llama-eval-callback",
         "llama-export-lora",
-        "llama-gbnf-validator",
         "llama-gen-docs",
         "llama-gguf",
         "llama-gguf-hash",
@@ -988,7 +1101,6 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
         "llama-perplexity",
         "llama-q8dot",
         "llama-quantize",
-        "llama-quantize-stats",
         "llama-qwen2vl-cli",
         "llama-retrieval",
         "llama-run",
@@ -1077,6 +1189,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
         return false;
+    } catch (std::exception & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        exit(1); // for other exceptions, we exit with status code 1
     }
 
     return true;
@@ -1377,13 +1492,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-f", "--file"}, "FNAME",
         "a file containing the prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
+            params.prompt = read_file(value);
             // store the external file name in params
             params.prompt_file = value;
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (!params.prompt.empty() && params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
@@ -1393,11 +1504,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-sysf", "--system-prompt-file"}, "FNAME",
         "a file containing the system prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            params.system_prompt = read_file(value);
             if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
                 params.system_prompt.pop_back();
             }
@@ -1822,15 +1929,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--grammar-file"}, "FNAME",
         "file to read grammar from",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sampling.grammar)
-            );
+            params.sampling.grammar = read_file(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1840,6 +1939,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"-jf", "--json-schema-file"}, "FILE",
+        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::string schema;
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(schema)
+            );
+            params.sampling.grammar = json_schema_to_grammar(json::parse(schema));
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--pooling"}, "{none,mean,cls,last,rank}",
         "pooling type for embeddings, use model default if unspecified",
@@ -2095,18 +2211,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
     add_opt(common_arg(
         {"--mmproj"}, "FILE",
-        "path to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        "path to a multimodal projector file. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.path = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples(mmproj_examples));
     add_opt(common_arg(
         {"--mmproj-url"}, "URL",
-        "URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        "URL to a multimodal projector file. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.url = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples(mmproj_examples));
+    add_opt(common_arg(
+        {"--no-mmproj"},
+        "explicitly disable multimodal projector, useful when using -hf",
+        [](common_params & params) {
+            params.no_mmproj = true;
+        }
+    ).set_examples(mmproj_examples));
+    add_opt(common_arg(
+        {"--no-mmproj-offload"},
+        "do not offload multimodal projector to GPU",
+        [](common_params & params) {
+            params.mmproj_use_gpu = false;
+        }
+    ).set_examples(mmproj_examples));
     add_opt(common_arg(
         {"--image"}, "FILE",
         "path to an image file. use with multimodal models. Specify multiple times for batching",
@@ -2381,6 +2511,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
         "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
+        "mmproj is also downloaded automatically if available. to disable, add --no-mmproj\n"
         "example: unsloth/phi-4-GGUF:q4_k_m\n"
         "(default: unused)",
         [](common_params & params, const std::string & value) {
@@ -2652,7 +2783,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
     add_opt(common_arg(
         {"--cache-reuse"}, "N",
-        string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse),
+        string_format(
+            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
+            "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
+        ),
         [](common_params & params, int value) {
             params.n_cache_reuse = value;
         }
@@ -2735,14 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
         ),
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.chat_template));
+            params.chat_template = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
     add_opt(common_arg(

common/arg.h

@@ -78,3 +78,12 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 
 // function to be used by test-arg-parser
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_has_curl();
+
+struct common_remote_params {
+    std::vector<std::string> headers;
+    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
+    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
+};
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);

common/common.h

@@ -342,6 +342,8 @@ struct common_params {
 
     // multimodal models (see examples/llava)
     struct common_params_model mmproj;
+    bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
     // embedding

common/json-schema-to-grammar.cpp

@@ -16,6 +16,9 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
+    if (max_items == 0) {
+        return "";
+    }
     if (min_items == 0 && max_items == 1) {
         return item_rule + "?";
     }

convert_hf_to_gguf.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
 from itertools import chain
+from transformers import AutoConfig
 
 import math
 import numpy as np
@@ -66,8 +67,6 @@ class ModelBase:
     part_names: list[str]
     is_safetensors: bool
     hparams: dict[str, Any]
-    block_count: int
-    tensor_map: gguf.TensorNameMap
     tensor_names: set[str] | None
     gguf_writer: gguf.GGUFWriter
     model_name: str | None
@@ -78,7 +77,11 @@ class ModelBase:
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
 
-    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
+    # subclasses should initialize this!
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
@@ -113,8 +116,6 @@ class ModelBase:
             if not self.is_safetensors:
                 self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = ModelBase.load_hparams(self.dir_model) if hparams is None else hparams
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         self.metadata_override = metadata_override
         self.model_name = model_name
@@ -417,11 +418,15 @@ class ModelBase:
 
     @staticmethod
     def load_hparams(dir_model: Path):
-        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-            hparams = json.load(f)
-            if "text_config" in hparams:
-                hparams = {**hparams, **hparams["text_config"]}
-            return hparams
+        try:
+            # for security reason, we don't allow loading remote code by default
+            # if a model need remote code, we will fallback to config.json
+            return AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
+        except Exception as e:
+            logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            logger.warning("Trying to load config.json instead")
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                return json.load(f)
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -450,6 +455,16 @@ class ModelBase:
 
 
 class TextModel(ModelBase):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        if "text_config" in self.hparams:
+            # move the text_config to the root level
+            self.hparams = {**self.hparams, **self.hparams["text_config"]}
+
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
     @classmethod
     def __init_subclass__(cls):
         # can't use an abstract property, because overriding it without type errors
@@ -772,6 +787,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
             # ref: https://huggingface.co/THUDM/glm-4-9b-hf
             res = "glm4"
+        if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
+            # ref: https://huggingface.co/mistral-community/pixtral-12b
+            res = "pixtral"
 
         if res is None:
             logger.warning("\n")
@@ -1061,6 +1079,8 @@ class TextModel(ModelBase):
 class VisionModel(ModelBase):
     model_arch = gguf.MODEL_ARCH.CLIP_VISION
     n_text_embd = 0
+    preprocessor_config: dict[str, Any]
+    global_config: dict[str, Any]
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -1068,31 +1088,43 @@ class VisionModel(ModelBase):
         if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION:
             raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")
 
-        # small hack to correct the number of layers
-        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, 128)
-        self.n_embd_text = self.find_hparam(["hidden_size", "n_embd"])
+        # get n_embd of the text model
+        text_config = {**self.hparams, **self.hparams["text_config"]}
+        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
         assert self.n_embd_text > 0, "n_embd not found in hparams"
 
         if "vision_config" not in self.hparams:
             raise ValueError("vision_config not found in hparams")
-        # move vision config to the top level
+        # move vision config to the top level, while preserving the original hparams in global_config
+        self.global_config = self.hparams
         self.hparams = self.hparams["vision_config"]
 
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"])
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count)
+
+        # load preprocessor config
+        with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
+            self.preprocessor_config = json.load(f)
+
     def set_type(self):
         self.gguf_writer.add_type(gguf.GGUFType.CLIP_VISION)
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.PROJECTION_DIM, self.n_embd_text)
-        self.gguf_writer.add_bool(gguf.Keys.ClipVision.HAS_VISION_ENCODER, True)
+        self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_vision_has_vision_encoder(True)
 
         # vision config
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.IMAGE_SIZE,           self.find_hparam(["image_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.PATCH_SIZE,           self.find_hparam(["patch_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.EMBEDDING_LENGTH,     self.find_hparam(["hidden_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.FEED_FORWARD_LENGTH,  self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.BLOCK_COUNT,          self.find_hparam(["num_hidden_layers"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.Attention.HEAD_COUNT, self.find_hparam(["num_attention_heads"]))
+        self.gguf_writer.add_vision_image_size(self.find_hparam(["image_size"]))
+        self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
+        self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
+        self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
+        self.gguf_writer.add_vision_block_count(self.block_count)
+        self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
+
+        # preprocessor config
+        self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
+        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
 
     def write_vocab(self):
         raise ValueError("VisionModel does not support vocab writing")
@@ -1703,7 +1735,13 @@ class StableLMModel(TextModel):
                 raise ValueError(f"Unprocessed norms: {norms}")
 
 
-@ModelBase.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+@ModelBase.register(
+    "LLaMAForCausalLM",
+    "LlamaForCausalLM",
+    "MistralForCausalLM",
+    "MixtralForCausalLM",
+    "VLlama3ForCausalLM",
+    "LlavaForConditionalGeneration")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
@@ -1770,6 +1808,17 @@ class LlamaModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
+        is_vision_tensor = "vision_tower" in name \
+            or "vision_model" in name \
+            or "model.connector" in name \
+            or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            return [] # skip vision tensors
+        elif name.startswith("model.text_model"):
+            name = name.replace("text_model.", "") # for SmolVLM
+        elif name.startswith("language_model."):
+            name = name.replace("language_model.", "") # for the rest
 
         if self.undo_permute:
             if name.endswith(("q_proj.weight", "q_proj.bias")):
@@ -1852,6 +1901,108 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register(
+    "LlavaForConditionalGeneration", # pixtral
+    "Mistral3ForConditionalGeneration", # mistral small 3.1
+)
+class LlavaVisionModel(VisionModel):
+    img_break_tok_id = -1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.hparams["model_type"] == "pixtral":
+            # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
+            self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
+            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
+            logger.info(f"Image break token id: {self.img_break_tok_id}")
+        else:
+            raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
+
+    def get_token_id(self, token: str) -> int:
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+            added_tokens_decoder = json.load(f)['added_tokens_decoder']
+            for id_, token_data in added_tokens_decoder.items():
+                if token_data["content"] == token:
+                    return int(id_)
+        raise ValueError(f"Token '{token}' not found in tokenizer config.")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if hparams["model_type"] == "pixtral":
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
+            self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
+
+            # hidden_act
+            if hparams["hidden_act"] == "silu":
+                self.gguf_writer.add_vision_use_silu(True)
+            elif hparams["hidden_act"] == "gelu":
+                self.gguf_writer.add_vision_use_gelu(True)
+            else:
+                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+
+            # spatial_merge_size
+            if "spatial_merge_size" in self.global_config:
+                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = n_head
+
+        if name.startswith("multi_modal_projector.") or name.startswith("vision_tower."):
+            # process vision tensors
+            if name.endswith(("q_proj.weight", "q_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+            if name.endswith(("k_proj.weight", "k_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+            return [(self.map_tensor_name(name), data_torch)]
+
+        if self.img_break_tok_id > 0 and "embed_tokens.weight" in name:
+            logger.info(f"Extracting [IMG_BREAK] token embedding from {name}")
+            # for pixtral model, we need to extract the [IMG_BREAK] token embedding
+            img_break_embd = data_torch[self.img_break_tok_id]
+            name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK]
+            return [(self.map_tensor_name(name), img_break_embd)]
+
+        return [] # skip other tensors
+
+
+@ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")
+class SmolVLMModel(VisionModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.hparams["model_type"] == "smolvlm_vision":
+            # fix for SmolVLM2, missing some keys in config.json
+            # default values are taken from transformers code
+            self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1152)
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
+            self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 3072)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.IDEFICS3)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
+        self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("scale_factor", 2))
+        self.gguf_writer.add_vision_use_gelu(True)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, new_name, n_dims  # unused
+        if ".embeddings." in name:
+            return gguf.GGMLQuantizationType.F32
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "vision_model" in name or "model.connector" in name
+
+        if is_vision_tensor:
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
+
 @ModelBase.register("Llama4ForConditionalGeneration")
 class Llama4Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.LLAMA4
@@ -2424,11 +2575,12 @@ class Qwen2VLModel(TextModel):
         except FileNotFoundError:
             self._set_vocab_gpt2()
 
-    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
-        for name, data in super().get_tensors():
-            if name.startswith("visual."):
-                continue
-            yield name, data
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        if name.startswith("visual."):
+            # skip visual tensors
+            return []
+        return [(self.map_tensor_name(name), data_torch)]
 
 
 @ModelBase.register("WavTokenizerDec")
@@ -3242,14 +3394,7 @@ class BertModel(TextModel):
 
         return [(self.map_tensor_name(name), data_torch)]
 
-
-@ModelBase.register("RobertaModel")
-class RobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
+    def _xlmroberta_tokenizer_init(self) -> None:
         # we need the pad_token_id to know how to chop down position_embd matrix
         if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
             self._position_offset = 1 + pad_token_id
@@ -3258,82 +3403,7 @@ class RobertaModel(BertModel):
         else:
             self._position_offset = None
 
-    def set_vocab(self):
-        """Support BPE tokenizers for roberta models"""
-        bpe_tok_path = self.dir_model / "tokenizer.json"
-        if bpe_tok_path.exists():
-            self._set_vocab_gpt2()
-            self.gguf_writer.add_add_bos_token(True)
-            self.gguf_writer.add_add_eos_token(True)
-
-            # we need this to validate the size of the token_type embeddings
-            # though currently we are passing all zeros to the token_type embeddings
-            # "Sequence A" or "Sequence B"
-            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
-
-        else:
-            return super().set_vocab()
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "roberta.", remove the prefix
-        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
-        if name.startswith("roberta."):
-            name = name[8:]
-
-        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
-        if name == "embeddings.position_embeddings.weight":
-            if self._position_offset is not None:
-                data_torch = data_torch[self._position_offset:,:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-
-@ModelBase.register("NomicBertModel")
-class NomicBertModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # the HF config claims n_ctx=8192, but it uses RoPE scaling
-        self.hparams["n_ctx"] = 2048
-
-        # SwigLU activation
-        assert self.hparams["activation_function"] == "swiglu"
-        # this doesn't do anything in the HF version
-        assert self.hparams["causal"] is False
-        # no bias tensors
-        assert self.hparams["qkv_proj_bias"] is False
-        assert self.hparams["mlp_fc1_bias"] is False
-        assert self.hparams["mlp_fc2_bias"] is False
-        # norm at end of layer
-        assert self.hparams["prenorm"] is False
-        # standard RoPE
-        assert self.hparams["rotary_emb_fraction"] == 1.0
-        assert self.hparams["rotary_emb_interleaved"] is False
-        assert self.hparams["rotary_emb_scale_base"] is None
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
-
-
-@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
-class XLMRobertaModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # we need the pad_token_id to know how to chop down position_embd matrix
-        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
-            self._position_offset = 1 + pad_token_id
-            if "max_position_embeddings" in self.hparams:
-                self.hparams["max_position_embeddings"] -= self._position_offset
-        else:
-            self._position_offset = None
-
-    def set_vocab(self):
+    def _xlmroberta_set_vocab(self) -> None:
         # to avoid TypeError: Descriptors cannot be created directly
         # exception when importing sentencepiece_model_pb2
         os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
@@ -3415,6 +3485,140 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
+
+@ModelBase.register("RobertaModel")
+class RobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
+
+    def set_vocab(self):
+        """Support BPE tokenizers for roberta models"""
+        bpe_tok_path = self.dir_model / "tokenizer.json"
+        if bpe_tok_path.exists():
+            self._set_vocab_gpt2()
+            self.gguf_writer.add_add_bos_token(True)
+            self.gguf_writer.add_add_eos_token(True)
+
+            # we need this to validate the size of the token_type embeddings
+            # though currently we are passing all zeros to the token_type embeddings
+            # "Sequence A" or "Sequence B"
+            self.gguf_writer.add_token_type_count(self.hparams.get("type_vocab_size", 1))
+
+        else:
+            return super().set_vocab()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("NomicBertModel")
+class NomicBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model)
+
+        self.is_moe = bool(hparams.get("moe_every_n_layers"))
+        self.model_arch = gguf.MODEL_ARCH.NOMIC_BERT_MOE if self.is_moe else gguf.MODEL_ARCH.NOMIC_BERT
+
+        super().__init__(dir_model, ftype, fname_out, hparams=hparams, **kwargs)
+
+        self._tokenizer_is_xlmroberta = self._is_tokenizer_xlmroberta()
+        if self._tokenizer_is_xlmroberta:
+            self._xlmroberta_tokenizer_init()
+
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048
+
+        assert self.hparams["activation_function"] == "gelu" if self.is_moe else "swiglu"
+
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors unless MoE
+        assert self.hparams["qkv_proj_bias"] == self.is_moe
+        assert self.hparams["mlp_fc1_bias"]  == self.is_moe
+        assert self.hparams["mlp_fc2_bias"]  == self.is_moe
+
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_vocab(self) -> None:
+        if self._tokenizer_is_xlmroberta:
+            return self._xlmroberta_set_vocab()
+        return super().set_vocab()
+
+    def modify_tensors(self, data_torch: torch.Tensor, name: str, bid: int | None) -> Iterable[tuple[str, torch.Tensor]]:
+        # If the tensor is an experts bias tensor, skip it by returning an empty list.
+        if "mlp.experts.bias" in name:
+            return []  # Explicitly return an empty list.
+
+        if "mlp.experts.mlp.w1" in name:
+            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            name += ".weight"
+
+        if "mlp.experts.mlp.w2" in name:
+            data_torch = data_torch.view(self.hparams["num_experts"], self.hparams["n_inner"], self.hparams["n_embd"])
+            data_torch = data_torch.transpose(1, 2)
+            name += ".weight"
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+        if self.is_moe:
+            self.gguf_writer.add_moe_every_n_layers(self.hparams["moe_every_n_layers"])
+            self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+            self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
+
+    def _is_tokenizer_xlmroberta(self) -> bool:
+        with open(self.dir_model / "tokenizer.json") as f:
+            tokenizer_json = json.load(f)
+        toktyp = tokenizer_json["model"]["type"]
+        if toktyp == "Unigram":
+            return True
+        if toktyp == "WordPiece":
+            return False
+        raise ValueError(f"unknown tokenizer: {toktyp}")
+
+
+@ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
+class XLMRobertaModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._xlmroberta_tokenizer_init()
+
+    def set_vocab(self):
+        self._xlmroberta_set_vocab()
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # if name starts with "roberta.", remove the prefix
         # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
@@ -3591,12 +3795,10 @@ class Gemma3VisionModel(VisionModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        self.gguf_writer.add_string(gguf.Keys.ClipVision.PROJECTOR_TYPE, "gemma3")
+        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.GEMMA3)
         # default values below are taken from HF tranformers code
-        self.gguf_writer.add_float32(gguf.Keys.ClipVision.Attention.LAYERNORM_EPS, hparams.get("layer_norm_eps", 1e-6))
-        self.gguf_writer.add_array(gguf.Keys.ClipVision.IMAGE_MEAN, [0.5, 0.5, 0.5])
-        self.gguf_writer.add_array(gguf.Keys.ClipVision.IMAGE_STD,  [0.5, 0.5, 0.5])
-        self.gguf_writer.add_bool (gguf.Keys.ClipVision.USE_GELU,   True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_vision_use_gelu(True)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         del bid, new_name, n_dims  # unused
@@ -3614,10 +3816,6 @@ class Gemma3VisionModel(VisionModel):
                 or name.startswith("multimodal_projector.") or name.startswith("vision_model."):
             # process vision tensors
             name = name.replace("_weight", ".weight")
-            if "fc1" in name:
-                name = name.replace("fc1", "fc2")
-            else:
-                name = name.replace("fc2", "fc1")
 
             # correct norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
             # the other norm values are part of SigLIP model, and they are already correct
@@ -5017,10 +5215,25 @@ class Glm4Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4
 
     def set_vocab(self):
-        self._set_vocab_gpt2()
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
+        rope_dim = self.hparams["head_dim"]
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "yarn":
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
@@ -5666,6 +5879,19 @@ def split_str_to_n_bytes(split_str: str) -> int:
     return n
 
 
+def get_model_architecture(dir_model: Path, model_type: ModelType, hparams: Any = None) -> str:
+    hparams = ModelBase.load_hparams(dir_model) if hparams is None else hparams
+    text_config = hparams.get("text_config", {})
+    vision_config = hparams.get("vision_config", {})
+    arch = hparams["architectures"][0]
+    # if "architectures" is found in the sub-config, use that instead
+    if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
+        arch = text_config["architectures"][0]
+    elif model_type == ModelType.VISION and vision_config.get("architectures") is not None:
+        arch = vision_config["architectures"][0]
+    return arch
+
+
 def main() -> None:
     args = parse_args()
 
@@ -5718,16 +5944,15 @@ def main() -> None:
 
     logger.info(f"Loading model: {dir_model.name}")
 
-    hparams = ModelBase.load_hparams(dir_model)
-
     if args.mmproj:
         if "mmproj" not in fname_out.name:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
-        model_architecture = hparams["architectures"][0]
         model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
+        model_architecture = get_model_architecture(dir_model, model_type)
+        logger.info(f"Model architecture: {model_architecture}")
         try:
             model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
         except NotImplementedError:

convert_hf_to_gguf_update.py

@@ -115,6 +115,7 @@ models = [
     {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
+    {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
 ]
 
 

docs/multimodal/gemma3.md

@@ -11,15 +11,15 @@ You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli
 
 # alternatively, install from brew (MacOS)
 brew install llama.cpp
 
 # run it
-llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
 
 # note: 1B model does not support vision
 ```
@@ -44,8 +44,8 @@ What you need:
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli
 
 # run it
-./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+./build/bin/llama-mtmd-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
 ```

examples/CMakeLists.txt

@@ -21,11 +21,6 @@ else()
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
 
-    if (NOT WIN32)
-        # disabled on Windows because it uses internal functions not exported with LLAMA_API
-        add_subdirectory(gbnf-validator)
-    endif()
-
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
@@ -58,10 +53,6 @@ else()
         add_subdirectory(convert-llama2c-to-ggml)
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
-        if (NOT WIN32)
-            # disabled on Windows because it uses internal functions not exported with LLAMA_API
-            add_subdirectory(quantize-stats)
-        endif()
         add_subdirectory(llava)
         if (GGML_RPC)
             add_subdirectory(rpc)

examples/embedding/embedding.cpp

@@ -89,6 +89,13 @@ int main(int argc, char ** argv) {
     common_init();
 
     params.embedding = true;
+
+    // utilize the full context
+    if (params.n_batch < params.n_ctx) {
+        LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);
+        params.n_batch = params.n_ctx;
+    }
+
     // For non-causal models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
 
@@ -134,7 +141,6 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch >= params.n_ctx);
 
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;

examples/gbnf-validator/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET llama-gbnf-validator)
-add_executable(${TARGET} gbnf-validator.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/json_schema_to_grammar.py

@@ -10,6 +10,9 @@ from typing import Any, List, Optional, Set, Tuple, Union
 
 def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
 
+    if max_items == 0:
+        return ""
+
     if min_items == 0 and max_items == 1:
         return f'{item_rule}?'
 

examples/llama-bench/README.md

@@ -28,6 +28,7 @@ options:
   -p, --n-prompt <n>                        (default: 512)
   -n, --n-gen <n>                           (default: 128)
   -pg <pp,tg>                               (default: )
+  -d, --n-depth <n>                         (default: 0)
   -b, --batch-size <n>                      (default: 2048)
   -ub, --ubatch-size <n>                    (default: 512)
   -ctk, --cache-type-k <t>                  (default: f16)
@@ -66,6 +67,8 @@ With the exception of `-r`, `-o` and `-v`, all options can be specified multiple
 
 Each test is repeated the number of times given by `-r`, and the results are averaged. The results are given in average tokens per second (t/s) and standard deviation. Some output formats (e.g. json) also include the individual results of each repetition.
 
+Using the `-d <n>` option, each test can be run at a specified context depth, prefilling the KV cache with `<n>` tokens.
+
 For a description of the other options, see the [main example](../main/README.md).
 
 Note:
@@ -148,6 +151,19 @@ $ ./llama-bench -ngl 10,20,30,31,32,33,34,35
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | pp 512     |   2400.01 ± 7.72 |
 | llama 7B mostly Q4_0           |   3.56 GiB |     6.74 B | CUDA       |  35 | tg 128     |    131.66 ± 0.49 |
 
+### Different prefilled context
+
+```
+$ ./llama-bench -d 0,512
+```
+
+| model                          |       size |     params | backend    | ngl |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | --------------: | -------------------: |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |           pp512 |      7340.20 ± 23.45 |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |           tg128 |        120.60 ± 0.59 |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |    pp512 @ d512 |      6425.91 ± 18.88 |
+| qwen2 7B Q4_K - Medium         |   4.36 GiB |     7.62 B | CUDA       |  99 |    tg128 @ d512 |        116.71 ± 0.60 |
+
 ## Output formats
 
 By default, llama-bench outputs the results in markdown format. The results can be output in other formats by using the `-o` option.
@@ -170,9 +186,9 @@ $ ./llama-bench -o csv
 ```
 
 ```csv
-build_commit,build_number,cuda,metal,gpu_blas,blas,cpu_info,gpu_info,model_filename,model_type,model_size,model_n_params,n_batch,n_threads,f16_kv,n_gpu_layers,main_gpu,mul_mat_q,tensor_split,n_prompt,n_gen,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
-"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","512","0","2023-09-23T12:09:01Z","212155977","732372","2413.341687","8.305961"
-"3469684","1275","1","0","0","1","1","13th Gen Intel(R) Core(TM) i9-13900K","NVIDIA GeForce RTX 3090 Ti","models/7B/ggml-model-q4_0.gguf","llama 7B mostly Q4_0","3825065984","6738415616","512","16","1","99","0","1","0.00","0","128","2023-09-23T12:09:02Z","969320879","2728399","132.052051","0.371342"
+build_commit,build_number,cpu_info,gpu_info,backends,model_filename,model_type,model_size,model_n_params,n_batch,n_ubatch,n_threads,cpu_mask,cpu_strict,poll,type_k,type_v,n_gpu_layers,split_mode,main_gpu,no_kv_offload,flash_attn,tensor_split,use_mmap,embeddings,n_prompt,n_gen,n_depth,test_time,avg_ns,stddev_ns,avg_ts,stddev_ts
+"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","512","0","0","2025-04-24T11:57:09Z","70285660","982040","7285.676949","100.064434"
+"8cf427ff","5163","AMD Ryzen 7 7800X3D 8-Core Processor","NVIDIA GeForce RTX 4080","CUDA","models/Qwen2.5-7B-Instruct-Q4_K_M.gguf","qwen2 7B Q4_K - Medium","4677120000","7615616512","2048","512","8","0x0","0","50","f16","f16","99","layer","0","0","0","0.00","1","0","0","128","0","2025-04-24T11:57:10Z","1067431600","3834831","119.915244","0.430617"
 ```
 
 ### JSON
@@ -184,64 +200,78 @@ $ ./llama-bench -o json
 ```json
 [
   {
-    "build_commit": "3469684",
-    "build_number": 1275,
-    "cuda": true,
-    "metal": false,
-    "gpu_blas": true,
-    "blas": true,
-    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
-    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
-    "model_filename": "models/7B/ggml-model-q4_0.gguf",
-    "model_type": "llama 7B mostly Q4_0",
-    "model_size": 3825065984,
-    "model_n_params": 6738415616,
-    "n_batch": 512,
-    "n_threads": 16,
-    "f16_kv": true,
+    "build_commit": "8cf427ff",
+    "build_number": 5163,
+    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
+    "gpu_info": "NVIDIA GeForce RTX 4080",
+    "backends": "CUDA",
+    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
+    "model_type": "qwen2 7B Q4_K - Medium",
+    "model_size": 4677120000,
+    "model_n_params": 7615616512,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 8,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
     "n_gpu_layers": 99,
+    "split_mode": "layer",
     "main_gpu": 0,
-    "mul_mat_q": true,
+    "no_kv_offload": false,
+    "flash_attn": false,
     "tensor_split": "0.00",
+    "use_mmap": true,
+    "embeddings": false,
     "n_prompt": 512,
     "n_gen": 0,
-    "test_time": "2023-09-23T12:09:57Z",
-    "avg_ns": 212365953,
-    "stddev_ns": 985423,
-    "avg_ts": 2410.974041,
-    "stddev_ts": 11.163766,
-    "samples_ns": [ 213837238, 211635853, 212328053, 211329715, 212698907 ],
-    "samples_ts": [ 2394.34, 2419.25, 2411.36, 2422.75, 2407.16 ]
+    "n_depth": 0,
+    "test_time": "2025-04-24T11:58:50Z",
+    "avg_ns": 72135640,
+    "stddev_ns": 1453752,
+    "avg_ts": 7100.002165,
+    "stddev_ts": 140.341520,
+    "samples_ns": [ 74601900, 71632900, 71745200, 71952700, 70745500 ],
+    "samples_ts": [ 6863.1, 7147.55, 7136.37, 7115.79, 7237.21 ]
   },
   {
-    "build_commit": "3469684",
-    "build_number": 1275,
-    "cuda": true,
-    "metal": false,
-    "gpu_blas": true,
-    "blas": true,
-    "cpu_info": "13th Gen Intel(R) Core(TM) i9-13900K",
-    "gpu_info": "NVIDIA GeForce RTX 3090 Ti",
-    "model_filename": "models/7B/ggml-model-q4_0.gguf",
-    "model_type": "llama 7B mostly Q4_0",
-    "model_size": 3825065984,
-    "model_n_params": 6738415616,
-    "n_batch": 512,
-    "n_threads": 16,
-    "f16_kv": true,
+    "build_commit": "8cf427ff",
+    "build_number": 5163,
+    "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor",
+    "gpu_info": "NVIDIA GeForce RTX 4080",
+    "backends": "CUDA",
+    "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf",
+    "model_type": "qwen2 7B Q4_K - Medium",
+    "model_size": 4677120000,
+    "model_n_params": 7615616512,
+    "n_batch": 2048,
+    "n_ubatch": 512,
+    "n_threads": 8,
+    "cpu_mask": "0x0",
+    "cpu_strict": false,
+    "poll": 50,
+    "type_k": "f16",
+    "type_v": "f16",
     "n_gpu_layers": 99,
+    "split_mode": "layer",
     "main_gpu": 0,
-    "mul_mat_q": true,
+    "no_kv_offload": false,
+    "flash_attn": false,
     "tensor_split": "0.00",
+    "use_mmap": true,
+    "embeddings": false,
     "n_prompt": 0,
     "n_gen": 128,
-    "test_time": "2023-09-23T12:09:59Z",
-    "avg_ns": 977425219,
-    "stddev_ns": 9268593,
-    "avg_ts": 130.965708,
-    "stddev_ts": 1.238924,
-    "samples_ns": [ 984472709, 974901233, 989474741, 970729355, 967548060 ],
-    "samples_ts": [ 130.019, 131.295, 129.362, 131.86, 132.293 ]
+    "n_depth": 0,
+    "test_time": "2025-04-24T11:58:51Z",
+    "avg_ns": 1076767880,
+    "stddev_ns": 9449585,
+    "avg_ts": 118.881588,
+    "stddev_ts": 1.041811,
+    "samples_ns": [ 1075361300, 1065089400, 1071761200, 1081934900, 1089692600 ],
+    "samples_ts": [ 119.03, 120.178, 119.43, 118.307, 117.464 ]
   }
 ]
 ```
@@ -254,8 +284,8 @@ $ ./llama-bench -o jsonl
 ```
 
 ```json lines
-{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":512,"n_gen":0,"test_time":"2023-09-23T12:09:57Z","avg_ns":212365953,"stddev_ns":985423,"avg_ts":2410.974041,"stddev_ts":11.163766,"samples_ns":[213837238,211635853,212328053,211329715,212698907],"samples_ts":[2394.34,2419.25,2411.36,2422.75,2407.16]}
-{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":0,"n_gen":128,"test_time":"2023-09-23T12:09:59Z","avg_ns":977425219,"stddev_ns":9268593,"avg_ts":130.965708,"stddev_ts":1.238924,"samples_ns":[984472709,974901233,989474741,970729355,967548060],"samples_ts":[130.019,131.295,129.362,131.86,132.293]}
+{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 512, "n_gen": 0, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 70497220, "stddev_ns": 883196, "avg_ts": 7263.609157, "stddev_ts": 90.940578, "samples_ns": [ 71551000, 71222800, 70364100, 69439100, 69909100 ],"samples_ts": [ 7155.74, 7188.71, 7276.44, 7373.37, 7323.8 ]}
+{"build_commit": "8cf427ff", "build_number": 5163, "cpu_info": "AMD Ryzen 7 7800X3D 8-Core Processor", "gpu_info": "NVIDIA GeForce RTX 4080", "backends": "CUDA", "model_filename": "models/Qwen2.5-7B-Instruct-Q4_K_M.gguf", "model_type": "qwen2 7B Q4_K - Medium", "model_size": 4677120000, "model_n_params": 7615616512, "n_batch": 2048, "n_ubatch": 512, "n_threads": 8, "cpu_mask": "0x0", "cpu_strict": false, "poll": 50, "type_k": "f16", "type_v": "f16", "n_gpu_layers": 99, "split_mode": "layer", "main_gpu": 0, "no_kv_offload": false, "flash_attn": false, "tensor_split": "0.00", "use_mmap": true, "embeddings": false, "n_prompt": 0, "n_gen": 128, "n_depth": 0, "test_time": "2025-04-24T11:59:33Z", "avg_ns": 1068078400, "stddev_ns": 6279455, "avg_ts": 119.844681, "stddev_ts": 0.699739, "samples_ns": [ 1066331700, 1064864900, 1079042600, 1063328400, 1066824400 ],"samples_ts": [ 120.038, 120.203, 118.624, 120.377, 119.982 ]}
 ```
 
 
@@ -271,25 +301,32 @@ $ ./llama-bench -o sql
 CREATE TABLE IF NOT EXISTS test (
   build_commit TEXT,
   build_number INTEGER,
-  cuda INTEGER,
-  metal INTEGER,
-  gpu_blas INTEGER,
-  blas INTEGER,
   cpu_info TEXT,
   gpu_info TEXT,
+  backends TEXT,
   model_filename TEXT,
   model_type TEXT,
   model_size INTEGER,
   model_n_params INTEGER,
   n_batch INTEGER,
+  n_ubatch INTEGER,
   n_threads INTEGER,
-  f16_kv INTEGER,
+  cpu_mask TEXT,
+  cpu_strict INTEGER,
+  poll INTEGER,
+  type_k TEXT,
+  type_v TEXT,
   n_gpu_layers INTEGER,
+  split_mode TEXT,
   main_gpu INTEGER,
-  mul_mat_q INTEGER,
+  no_kv_offload INTEGER,
+  flash_attn INTEGER,
   tensor_split TEXT,
+  use_mmap INTEGER,
+  embeddings INTEGER,
   n_prompt INTEGER,
   n_gen INTEGER,
+  n_depth INTEGER,
   test_time TEXT,
   avg_ns INTEGER,
   stddev_ns INTEGER,
@@ -297,6 +334,6 @@ CREATE TABLE IF NOT EXISTS test (
   stddev_ts REAL
 );
 
-INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '512', '0', '2023-09-23T12:10:30Z', '212693772', '743623', '2407.240204', '8.409634');
-INSERT INTO test (build_commit, build_number, cuda, metal, gpu_blas, blas, cpu_info, gpu_info, model_filename, model_type, model_size, model_n_params, n_batch, n_threads, f16_kv, n_gpu_layers, main_gpu, mul_mat_q, tensor_split, n_prompt, n_gen, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('3469684', '1275', '1', '0', '0', '1', '1', '13th Gen Intel(R) Core(TM) i9-13900K', 'NVIDIA GeForce RTX 3090 Ti', 'models/7B/ggml-model-q4_0.gguf', 'llama 7B mostly Q4_0', '3825065984', '6738415616', '512', '16', '1', '99', '0', '1', '0.00', '0', '128', '2023-09-23T12:10:31Z', '977925003', '4037361', '130.891159', '0.537692');
+INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '512', '0', '0', '2025-04-24T12:00:08Z', '69905000', '519516', '7324.546977', '54.032613');
+INSERT INTO test (build_commit, build_number, cpu_info, gpu_info, backends, model_filename, model_type, model_size, model_n_params, n_batch, n_ubatch, n_threads, cpu_mask, cpu_strict, poll, type_k, type_v, n_gpu_layers, split_mode, main_gpu, no_kv_offload, flash_attn, tensor_split, use_mmap, embeddings, n_prompt, n_gen, n_depth, test_time, avg_ns, stddev_ns, avg_ts, stddev_ts) VALUES ('8cf427ff', '5163', 'AMD Ryzen 7 7800X3D 8-Core Processor', 'NVIDIA GeForce RTX 4080', 'CUDA', 'models/Qwen2.5-7B-Instruct-Q4_K_M.gguf', 'qwen2 7B Q4_K - Medium', '4677120000', '7615616512', '2048', '512', '8', '0x0', '0', '50', 'f16', 'f16', '99', 'layer', '0', '0', '0', '0.00', '1', '0', '0', '128', '0', '2025-04-24T12:00:09Z', '1063608780', '4464130', '120.346696', '0.504647');
 ```

examples/llama-bench/llama-bench.cpp

@@ -36,6 +36,46 @@ static uint64_t get_time_ns() {
     return std::chrono::nanoseconds(clock::now().time_since_epoch()).count();
 }
 
+static bool tensor_buft_override_equal(const llama_model_tensor_buft_override& a, const llama_model_tensor_buft_override& b) {
+    if (a.pattern != b.pattern) {
+        // cString comparison that may be null
+        if (a.pattern == nullptr || b.pattern == nullptr) {
+            return false;
+        }
+        if (strcmp(a.pattern, b.pattern) != 0) {
+            return false;
+        }
+    }
+    if (a.buft != b.buft) {
+        return false;
+    }
+    return true;
+}
+
+static bool vec_tensor_buft_override_equal(const std::vector<llama_model_tensor_buft_override>& a, const std::vector<llama_model_tensor_buft_override>& b) {
+    if (a.size() != b.size()) {
+        return false;
+    }
+    for (size_t i = 0; i < a.size(); i++) {
+        if (!tensor_buft_override_equal(a[i], b[i])) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool vec_vec_tensor_buft_override_equal(const std::vector<std::vector<llama_model_tensor_buft_override>>& a, const std::vector<std::vector<llama_model_tensor_buft_override>>& b) {
+    if (a.size() != b.size()) {
+        return false;
+    }
+    for (size_t i = 0; i < a.size(); i++) {
+        if (!vec_tensor_buft_override_equal(a[i], b[i])) {
+            return false;
+        }
+    }
+    return true;
+}
+
 template <class T> static std::string join(const std::vector<T> & values, const std::string & delim) {
     std::ostringstream str;
     for (size_t i = 0; i < values.size(); i++) {
@@ -160,6 +200,7 @@ struct cmd_params {
     std::vector<int>                 n_prompt;
     std::vector<int>                 n_gen;
     std::vector<std::pair<int, int>> n_pg;
+    std::vector<int>                 n_depth;
     std::vector<int>                 n_batch;
     std::vector<int>                 n_ubatch;
     std::vector<ggml_type>           type_k;
@@ -175,6 +216,7 @@ struct cmd_params {
     std::vector<bool>                no_kv_offload;
     std::vector<bool>                flash_attn;
     std::vector<std::vector<float>>  tensor_split;
+    std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides;
     std::vector<bool>                use_mmap;
     std::vector<bool>                embeddings;
     ggml_numa_strategy               numa;
@@ -192,6 +234,7 @@ static const cmd_params cmd_params_defaults = {
     /* n_prompt             */ { 512 },
     /* n_gen                */ { 128 },
     /* n_pg                 */ {},
+    /* n_depth              */ { 0 },
     /* n_batch              */ { 2048 },
     /* n_ubatch             */ { 512 },
     /* type_k               */ { GGML_TYPE_F16 },
@@ -207,6 +250,7 @@ static const cmd_params cmd_params_defaults = {
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { false },
     /* tensor_split         */ { std::vector<float>(llama_max_devices(), 0.0f) },
+    /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{{nullptr,nullptr}} },
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
@@ -230,6 +274,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -n, --n-gen <n>                           (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -pg <pp,tg>                               (default: %s)\n",
            join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
+    printf("  -d, --n-depth <n>                         (default: %s)\n", join(cmd_params_defaults.n_depth, ",").c_str());
     printf("  -b, --batch-size <n>                      (default: %s)\n",
            join(cmd_params_defaults.n_batch, ",").c_str());
     printf("  -ub, --ubatch-size <n>                    (default: %s)\n",
@@ -265,6 +310,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -embd, --embeddings <0|1>                 (default: %s)\n",
            join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
+    printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;... (default: disabled)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
     printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
     printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
@@ -366,6 +412,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.n_pg.push_back({ std::stoi(p[0]), std::stoi(p[1]) });
+        } else if (arg == "-d" || arg == "--n-depth") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.n_depth.insert(params.n_depth.end(), p.begin(), p.end());
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -557,6 +610,87 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 params.tensor_split.push_back(tensor_split);
             }
+        } else if (arg == "-ot" || arg == "--override-tensor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto value = argv[i];
+            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+            if (buft_list.empty()) {
+                // enumerate all the devices and add their buffer types to the list
+                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                    auto * dev = ggml_backend_dev_get(i);
+                    auto * buft = ggml_backend_dev_buffer_type(dev);
+                    if (buft) {
+                        buft_list[ggml_backend_buft_name(buft)] = buft;
+                    }
+                }
+            }
+            auto override_group_span_len = std::strcspn(value, ",");
+            bool last_group = false;
+            do {
+                if (override_group_span_len == 0) {
+                    // Adds an empty override-tensors for an empty span
+                    params.tensor_buft_overrides.push_back({{}});
+                    if (value[override_group_span_len] == '\0') {
+                        value = &value[override_group_span_len];
+                        last_group = true;
+                    } else {
+                        value = &value[override_group_span_len + 1];
+                        override_group_span_len = std::strcspn(value, ",");
+                    }
+                    continue;
+                }
+                // Stamps null terminators into the argv
+                // value for this option to avoid the
+                // memory leak present in the implementation
+                // over in arg.cpp. Acceptable because we
+                // only parse these args once in this program.
+                auto override_group = value;
+                if (value[override_group_span_len] == '\0') {
+                    value = &value[override_group_span_len];
+                    last_group = true;
+                } else {
+                    value[override_group_span_len] = '\0';
+                    value = &value[override_group_span_len + 1];
+                }
+                std::vector<llama_model_tensor_buft_override> group_tensor_buft_overrides{};
+                auto override_span_len = std::strcspn(override_group, ";");
+                while (override_span_len > 0) {
+                    auto override = override_group;
+                    if (override_group[override_span_len] != '\0') {
+                        override_group[override_span_len] = '\0';
+                        override_group = &override_group[override_span_len + 1];
+                    } else {
+                        override_group = &override_group[override_span_len];
+                    }
+                    auto tensor_name_span_len = std::strcspn(override, "=");
+                    if (tensor_name_span_len >= override_span_len) {
+                        invalid_param = true;
+                        break;
+                    }
+                    override[tensor_name_span_len] = '\0';
+                    auto tensor_name = override;
+                    auto buffer_type = &override[tensor_name_span_len + 1];
+                    if (buft_list.find(buffer_type) == buft_list.end()) {
+                        printf("Available buffer types:\n");
+                        for (const auto & it : buft_list) {
+                            printf("  %s\n", ggml_backend_buft_name(it.second));
+                        }
+                        invalid_param = true;
+                        break;
+                    }
+                    group_tensor_buft_overrides.push_back({tensor_name, buft_list.at(buffer_type)});
+                    override_span_len = std::strcspn(override_group, ";");
+                }
+                if (invalid_param) {
+                    break;
+                }
+                group_tensor_buft_overrides.push_back({nullptr,nullptr});
+                params.tensor_buft_overrides.push_back(group_tensor_buft_overrides);
+                override_group_span_len = std::strcspn(value, ",");
+            } while (!last_group);
         } else if (arg == "-r" || arg == "--repetitions") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -615,6 +749,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_pg.empty()) {
         params.n_pg = cmd_params_defaults.n_pg;
     }
+    if (params.n_depth.empty()) {
+        params.n_depth = cmd_params_defaults.n_depth;
+    }
     if (params.n_batch.empty()) {
         params.n_batch = cmd_params_defaults.n_batch;
     }
@@ -648,6 +785,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.tensor_split.empty()) {
         params.tensor_split = cmd_params_defaults.tensor_split;
     }
+    if (params.tensor_buft_overrides.empty()) {
+        params.tensor_buft_overrides = cmd_params_defaults.tensor_buft_overrides;
+    }
     if (params.use_mmap.empty()) {
         params.use_mmap = cmd_params_defaults.use_mmap;
     }
@@ -674,6 +814,7 @@ struct cmd_params_instance {
     std::string        model;
     int                n_prompt;
     int                n_gen;
+    int                n_depth;
     int                n_batch;
     int                n_ubatch;
     ggml_type          type_k;
@@ -689,6 +830,7 @@ struct cmd_params_instance {
     bool               no_kv_offload;
     bool               flash_attn;
     std::vector<float> tensor_split;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool               use_mmap;
     bool               embeddings;
 
@@ -733,19 +875,26 @@ struct cmd_params_instance {
         mparams.tensor_split = tensor_split.data();
         mparams.use_mmap     = use_mmap;
 
+        if (tensor_buft_overrides.empty()) {
+            mparams.tensor_buft_overrides = nullptr;
+        } else {
+            GGML_ASSERT(tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
+            mparams.tensor_buft_overrides = tensor_buft_overrides.data();
+        }
+
         return mparams;
     }
 
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers_str == other.rpc_servers_str &&
                split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap &&
-               tensor_split == other.tensor_split;
+               tensor_split == other.tensor_split && vec_tensor_buft_override_equal(tensor_buft_overrides, other.tensor_buft_overrides);
     }
 
     llama_context_params to_llama_cparams() const {
         llama_context_params cparams = llama_context_default_params();
 
-        cparams.n_ctx       = n_prompt + n_gen;
+        cparams.n_ctx       = n_prompt + n_gen + n_depth;
         cparams.n_batch     = n_batch;
         cparams.n_ubatch    = n_ubatch;
         cparams.type_k      = type_k;
@@ -769,6 +918,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
+    for (const auto & ot : params.tensor_buft_overrides)
     for (const auto & mmp : params.use_mmap)
     for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
@@ -780,6 +930,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & nt : params.n_threads)
     for (const auto & cm : params.cpu_mask)
     for (const auto & cs : params.cpu_strict)
+    for (const auto & nd : params.n_depth)
     for (const auto & pl : params.poll) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
@@ -789,6 +940,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .model        = */ m,
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
+                /* .n_depth      = */ nd,
                 /* .n_batch      = */ nb,
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
@@ -804,6 +956,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
+                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -818,6 +971,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .model        = */ m,
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
+                /* .n_depth      = */ nd,
                 /* .n_batch      = */ nb,
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
@@ -833,6 +987,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
+                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -847,6 +1002,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .model        = */ m,
                 /* .n_prompt     = */ n_pg.first,
                 /* .n_gen        = */ n_pg.second,
+                /* .n_depth      = */ nd,
                 /* .n_batch      = */ nb,
                 /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
@@ -862,6 +1018,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
+                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -896,10 +1053,12 @@ struct test {
     bool                     no_kv_offload;
     bool                     flash_attn;
     std::vector<float>       tensor_split;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool                     use_mmap;
     bool                     embeddings;
     int                      n_prompt;
     int                      n_gen;
+    int                      n_depth;
     std::string              test_time;
     std::vector<uint64_t>    samples_ns;
 
@@ -927,10 +1086,12 @@ struct test {
         no_kv_offload  = inst.no_kv_offload;
         flash_attn     = inst.flash_attn;
         tensor_split   = inst.tensor_split;
+        tensor_buft_overrides = inst.tensor_buft_overrides;
         use_mmap       = inst.use_mmap;
         embeddings     = inst.embeddings;
         n_prompt       = inst.n_prompt;
         n_gen          = inst.n_gen;
+        n_depth        = inst.n_depth;
         // RFC 3339 date-time format
         time_t t       = time(NULL);
         std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
@@ -972,9 +1133,9 @@ struct test {
             "build_commit", "build_number", "cpu_info",       "gpu_info",   "backends",     "model_filename",
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
-            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "use_mmap",
-            "embeddings",   "n_prompt",     "n_gen",          "test_time",  "avg_ns",       "stddev_ns",
-            "avg_ts",       "stddev_ts",
+            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
+            "use_mmap",     "embeddings",   "n_prompt",       "n_gen",      "n_depth",      "test_time",
+            "avg_ns",       "stddev_ns",    "avg_ts",         "stddev_ts",
         };
         return fields;
     }
@@ -984,8 +1145,8 @@ struct test {
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" || field == "n_threads" ||
             field == "poll" || field == "model_size" || field == "model_n_params" || field == "n_gpu_layers" ||
-            field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "avg_ns" ||
-            field == "stddev_ns") {
+            field == "main_gpu" || field == "n_prompt" || field == "n_gen" || field == "n_depth" ||
+            field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
         if (field == "f16_kv" || field == "no_kv_offload" || field == "cpu_strict" || field == "flash_attn" ||
@@ -1000,6 +1161,7 @@ struct test {
 
     std::vector<std::string> get_values() const {
         std::string tensor_split_str;
+        std::string tensor_buft_overrides_str;
         int         max_nonzero = 0;
         for (size_t i = 0; i < llama_max_devices(); i++) {
             if (tensor_split[i] > 0) {
@@ -1014,6 +1176,26 @@ struct test {
                 tensor_split_str += "/";
             }
         }
+        if (tensor_buft_overrides.size() == 1) {
+            // Last element of tensor_buft_overrides is always a null pattern
+            // so if it is only one element long, it must be a null pattern.
+            GGML_ASSERT(tensor_buft_overrides[0].pattern == nullptr);
+            tensor_buft_overrides_str += "none";
+        } else {
+            for (size_t i = 0; i < tensor_buft_overrides.size()-1; i++) {
+                // Last element of tensor_buft_overrides is always a null pattern
+                if (tensor_buft_overrides[i].pattern == nullptr) {
+                    tensor_buft_overrides_str += "none";
+                } else {
+                    tensor_buft_overrides_str += tensor_buft_overrides[i].pattern;
+                    tensor_buft_overrides_str += "=";
+                    tensor_buft_overrides_str += ggml_backend_buft_name(tensor_buft_overrides[i].buft);
+                }
+                if (i + 2 < tensor_buft_overrides.size()) {
+                    tensor_buft_overrides_str += ";";
+                }
+            }
+        }
         std::vector<std::string> values = { build_commit,
                                             std::to_string(build_number),
                                             cpu_info,
@@ -1037,10 +1219,12 @@ struct test {
                                             std::to_string(no_kv_offload),
                                             std::to_string(flash_attn),
                                             tensor_split_str,
+                                            tensor_buft_overrides_str,
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
                                             std::to_string(n_prompt),
                                             std::to_string(n_gen),
+                                            std::to_string(n_depth),
                                             test_time,
                                             std::to_string(avg_ns()),
                                             std::to_string(stdev_ns()),
@@ -1218,7 +1402,7 @@ struct markdown_printer : public printer {
             return 4;
         }
         if (field == "test") {
-            return 13;
+            return 15;
         }
 
         int width = std::max((int) field.length(), 10);
@@ -1254,6 +1438,9 @@ struct markdown_printer : public printer {
         if (field == "tensor_split") {
             return "ts";
         }
+        if (field == "tensor_buft_overrides") {
+            return "ot";
+        }
         return field;
     }
 
@@ -1307,6 +1494,9 @@ struct markdown_printer : public printer {
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.emplace_back("tensor_split");
         }
+        if (params.tensor_buft_overrides.size() > 1 || !vec_vec_tensor_buft_override_equal(params.tensor_buft_overrides, cmd_params_defaults.tensor_buft_overrides)) {
+            fields.emplace_back("tensor_buft_overrides");
+        }
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }
@@ -1362,6 +1552,10 @@ struct markdown_printer : public printer {
                 } else {
                     snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen);
                 }
+                if (t.n_depth > 0) {
+                    int len = strlen(buf);
+                    snprintf(buf + len, sizeof(buf) - len, " @ d%d", t.n_depth);
+                }
                 value = buf;
             } else if (field == "t/s") {
                 snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.avg_ts(), t.stdev_ts());
@@ -1620,6 +1814,14 @@ int main(int argc, char ** argv) {
         for (int i = 0; i < params.reps; i++) {
             llama_kv_self_clear(ctx);
 
+            if (t.n_depth > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count,
+                            i + 1, params.reps);
+                }
+                test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads);
+            }
+
             uint64_t t_start = get_time_ns();
 
             if (t.n_prompt > 0) {

examples/llava/CMakeLists.txt

@@ -64,13 +64,7 @@ endif()
 add_executable(llama-llava-cli    deprecation-warning.cpp)
 add_executable(llama-gemma3-cli   deprecation-warning.cpp)
 add_executable(llama-minicpmv-cli deprecation-warning.cpp)
-
-set(TARGET llama-qwen2vl-cli)
-add_executable(${TARGET} qwen2vl-cli.cpp)
-set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
+add_executable(llama-qwen2vl-cli  deprecation-warning.cpp)
 
 set(TARGET llama-mtmd-cli)
 add_executable(${TARGET} mtmd-cli.cpp)

examples/llava/README.md

@@ -14,6 +14,31 @@ The naming and structure related to multimodal support have evolved, which might
 - [#12849](https://github.com/ggml-org/llama.cpp/pull/12849): `libmtmd` was introduced as a replacement for `llava.cpp`. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs.
 - [#13012](https://github.com/ggml-org/llama.cpp/pull/13012): `mtmd-cli` was added, consolidating the various model-specific CLIs into a single tool powered by `libmtmd`.
 
+## Pre-quantized models
+
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
+
+```sh
+# Gemma 3
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+llama-mtmd-cli -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF --chat-template mistral-v7
+```
+
 ## How it works and what is `mmproj`?
 
 Multimodal support in `llama.cpp` works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
@@ -45,3 +70,10 @@ Multimodal projector (`mmproj`) files are specific to each model architecture. P
 - [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
 - [IBM Granite Vision](../../docs/multimodal/granitevision.md)
 - [Google Gemma 3](../../docs/multimodal/gemma3.md)
+
+For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
+- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
+- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
+- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)

examples/llava/clip-impl.h

@@ -2,8 +2,6 @@
 #include "gguf.h"
 #include "clip.h"
 
-#include "clip.h"
-
 #include <climits>
 #include <cstdarg>
 #include <string>
@@ -17,33 +15,32 @@
 #define KEY_FTYPE               "general.file_type"
 #define KEY_NAME                "general.name"
 #define KEY_DESCRIPTION         "general.description"
-#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
-#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
-#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
-#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
-#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
-#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_USE_SILU            "clip.use_silu"
-#define KEY_N_EMBD              "clip.%s.embedding_length"
-#define KEY_N_FF                "clip.%s.feed_forward_length"
-#define KEY_N_BLOCK             "clip.%s.block_count"
-#define KEY_N_HEAD              "clip.%s.attention.head_count"
-#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM            "clip.%s.projection_dim"
-#define KEY_TOKENS              "tokenizer.ggml.tokens"
-#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_N_EMBD              "clip.vision.embedding_length"
+#define KEY_N_FF                "clip.vision.feed_forward_length"
+#define KEY_N_BLOCK             "clip.vision.block_count"
+#define KEY_N_HEAD              "clip.vision.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.vision.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.vision.projection_dim"
 #define KEY_IMAGE_SIZE          "clip.vision.image_size"
 #define KEY_PATCH_SIZE          "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
-#define KEY_PROJ_TYPE           "clip.projector_type"
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+#define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+#define KEY_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
+
+#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
+#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
 #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+#define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
+#define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 
 
 //
@@ -60,7 +57,9 @@
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
 #define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_LN_1            "%s.blk.%d.ln1.%s"
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
@@ -70,8 +69,12 @@
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_NORM     "mm.input_norm.weight"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
+#define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
+#define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -87,18 +90,19 @@
 #define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
-#define TN_GLM_BOI_W            "adapter.boi"
-#define TN_GLM_EOI_W            "adapter.eoi"
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
-    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_MINICPMV,
     PROJECTOR_TYPE_GLM_EDGE,
-    PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_QWEN2VL,
     PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_IDEFICS3,
+    PROJECTOR_TYPE_PIXTRAL,
+    PROJECTOR_TYPE_QWEN25VL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -106,10 +110,13 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP,       "mlp" },
     { PROJECTOR_TYPE_LDP,       "ldp" },
     { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
-    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_MINICPMV,  "resampler"},
     { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
-    { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
+    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

examples/llava/clip.h

@@ -47,7 +47,7 @@ CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_par
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
 CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
+CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
 
 CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
 CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
@@ -59,9 +59,20 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
 CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
 
-CLIP_API int clip_n_patches        (const struct clip_ctx * ctx);
-CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
-CLIP_API int clip_n_mmproj_embd    (const struct clip_ctx * ctx);
+GGML_DEPRECATED(CLIP_API int clip_n_patches(const struct clip_ctx * ctx),
+    "use clip_n_output_tokens instead");
+GGML_DEPRECATED(CLIP_API int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img),
+    "use clip_n_output_tokens instead");
+
+CLIP_API int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// for M-RoPE, this will be the number of token positions in X and Y directions
+// for other models, X will be the total number of tokens and Y will be 1
+CLIP_API int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+CLIP_API int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
+
+// this should be equal to the embedding dimension of the text model
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
 CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
@@ -114,8 +125,6 @@ CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
 
-CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
-
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
 

examples/llava/llava.cpp

@@ -112,7 +112,7 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<
 }
 
 // Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
-static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out, clip_image_f32 * img_input) {
     struct {
         struct ggml_context * ctx;
     } model;
@@ -175,7 +175,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     model.ctx = ggml_init(params);
 
-    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_output_tokens(ctx_clip, img_input), num_images - 1); // example: 4096 x 576 x 4
     // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
     // fill it with the image embeddings, ignoring the base
     for (size_t i = 1; i < num_images; i++) {
@@ -214,8 +214,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
     // append without newline tokens (default behavior in llava_arch when not using unpad ):
-    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
-    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+    memcpy(image_embd_out + clip_n_output_tokens(ctx_clip, img_input) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_output_tokens(ctx_clip, img_input));
 
     // Debug: Test single segments
     // Current findings: sending base image, sending a segment embedding all works similar to python
@@ -313,7 +313,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
                 image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
                 image_embd_v[i],
                 clip_embd_nbytes_by_img(ctx_clip, nx, ny));
-            n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
+            n_img_pos_out += clip_n_output_tokens(ctx_clip, img_res);
         }
         *n_img_pos = n_img_pos_out;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -342,8 +342,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     }
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
-        *n_img_pos = clip_n_patches(ctx_clip);
         clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
+        *n_img_pos = clip_n_output_tokens(ctx_clip, img_res);
         bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
         if (!encoded) {
             LOG_ERR("Unable to encode image\n");
@@ -381,7 +381,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
 
         int n_img_pos_out;
-        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        clip_image_f32 * img_input = clip_image_f32_get_img(img_res_v.get(), 0);
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out, img_input);
         *n_img_pos = n_img_pos_out;
 
         for (size_t i = 0; i < image_embd_v.size(); i++) {

examples/llava/mtmd-cli.cpp

@@ -24,7 +24,9 @@
 #include <signal.h>
 #endif
 
-static bool g_is_generating = false;
+// volatile, because of signal being an interrupt
+static volatile bool g_is_generating = false;
+static volatile bool g_is_interrupted = false;
 
 /**
  * Please note that this is NOT a production-ready stuff.
@@ -38,7 +40,8 @@ static void show_additional_info(int /*argc*/, char ** argv) {
         "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
         "  -m and --mmproj are required\n"
         "  -hf user/repo can replace both -m and --mmproj in most cases\n"
-        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
+        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n"
+        "  to disable using GPU for mmproj model, add --no-mmproj-offload\n",
         argv[0]
     );
 }
@@ -50,8 +53,10 @@ static void sigint_handler(int signo) {
             g_is_generating = false;
         } else {
             console::cleanup();
-            LOG("\nInterrupted by user\n");
-            _exit(130);
+            if (g_is_interrupted) {
+                _exit(1);
+            }
+            g_is_interrupted = true;
         }
     }
 }
@@ -67,6 +72,8 @@ struct mtmd_cli_context {
     llama_batch         batch;
     int                 n_batch;
 
+    std::vector<mtmd_bitmap> bitmaps;
+
     // note: we know that gemma3 template is "linear", meaning each turn is completely separated to another
     // so here we don't need to keep track of chat history
     common_chat_templates_ptr tmpls;
@@ -89,6 +96,7 @@ struct mtmd_cli_context {
             LOG_ERR("Model does not have chat template.\n");
             LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
             LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            LOG_ERR("  For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
             exit(1);
         }
 
@@ -108,10 +116,10 @@ struct mtmd_cli_context {
     void init_vision_context(common_params & params) {
         const char * clip_path = params.mmproj.path.c_str();
         ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{
-            /* use_gpu */   true,
+            /* use_gpu */   params.mmproj_use_gpu,
             /* timings */   true,
             /* n_threads */ params.cpuparams.n_threads,
-            /* verbosity */ GGML_LOG_LEVEL_INFO,
+            /* verbosity */ params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO,
         }));
         if (!ctx_vision.get()) {
             LOG_ERR("Failed to load vision model from %s\n", clip_path);
@@ -129,46 +137,22 @@ struct mtmd_cli_context {
             antiprompt_tokens.begin()
         );
     }
-};
 
-struct decode_embd_batch {
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-        for (int i = 0; i < n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
+    bool load_image(const std::string & fname) {
+        mtmd_bitmap bitmap;
+        if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
+            return false;
         }
+        bitmaps.push_back(std::move(bitmap));
+        return true;
     }
 };
 
 static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
     llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
-        if (i > n_predict || !g_is_generating) {
-            printf("\n");
+        if (i > n_predict || !g_is_generating || g_is_interrupted) {
+            LOG("\n");
             break;
         }
 
@@ -177,13 +161,18 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
         common_sampler_accept(smpl, token_id, true);
 
         if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
-            printf("\n");
+            LOG("\n");
             break; // end of generation
         }
 
-        printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
+        LOG("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
         fflush(stdout);
 
+        if (g_is_interrupted) {
+            LOG("\n");
+            break;
+        }
+
         // eval the token
         common_batch_clear(ctx.batch);
         common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
@@ -195,9 +184,7 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
     return 0;
 }
 
-static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
-    std::vector<mtmd_bitmap> bitmaps;
-
+static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, bool add_bos = false) {
     common_chat_templates_inputs tmpl_inputs;
     tmpl_inputs.messages = {msg};
     tmpl_inputs.add_generation_prompt = true;
@@ -205,32 +192,30 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vect
     auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs);
     LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str());
 
-    for (auto & fname : images_fname) {
-        mtmd_bitmap bitmap;
-        if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) {
-            LOG_ERR("Unable to load image %s\n", fname.c_str());
-            return 2; // image not found
-        }
-        bitmaps.push_back(std::move(bitmap));
-    }
-
     mtmd_input_text text;
     text.text          = formatted_chat.prompt;
     text.add_special   = add_bos;
     text.parse_special = true;
     mtmd_input_chunks chunks;
-    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
+
+    if (g_is_interrupted) return 0;
+
+    int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, ctx.bitmaps);
     if (res != 0) {
         LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
         return 1;
     }
 
+    ctx.bitmaps.clear();
+
     if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks, ctx.n_past, 0, ctx.n_batch)) {
         LOG_ERR("Unable to eval prompt\n");
         return 1;
     }
 
-    ctx.n_past += mtmd_helper_get_n_tokens(chunks);
+    ctx.n_past += mtmd_helper_get_n_pos(chunks);
+
+    LOG("\n");
 
     return 0;
 }
@@ -249,11 +234,12 @@ int main(int argc, char ** argv) {
 
     if (params.mmproj.path.empty()) {
         show_additional_info(argc, argv);
+        LOG_ERR("ERR: Missing --mmproj argument\n");
         return 1;
     }
 
     mtmd_cli_context ctx(params);
-    printf("%s: %s\n", __func__, params.model.path.c_str());
+    LOG("%s: loading model: %s\n", __func__, params.model.path.c_str());
 
     bool is_single_turn = !params.prompt.empty() && !params.image.empty();
 
@@ -276,6 +262,8 @@ int main(int argc, char ** argv) {
 #endif
     }
 
+    if (g_is_interrupted) return 130;
+
     if (is_single_turn) {
         g_is_generating = true;
         if (params.prompt.find("<__image__>") == std::string::npos) {
@@ -284,10 +272,15 @@ int main(int argc, char ** argv) {
         common_chat_msg msg;
         msg.role = "user";
         msg.content = params.prompt;
-        if (eval_message(ctx, msg, params.image, true)) {
+        for (const auto & image : params.image) {
+            if (!ctx.load_image(image)) {
+                return 1; // error is already printed by libmtmd
+            }
+        }
+        if (eval_message(ctx, msg, true)) {
             return 1;
         }
-        if (generate_response(ctx, smpl, n_predict)) {
+        if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
             return 1;
         }
 
@@ -299,15 +292,15 @@ int main(int argc, char ** argv) {
         LOG("\n");
 
         bool is_first_msg = true;
-        std::vector<std::string> images_fname;
         std::string content;
 
-        while (true) {
+        while (!g_is_interrupted) {
             g_is_generating = false;
             LOG("\n> ");
             console::set_display(console::user_input);
             std::string line;
             console::readline(line, false);
+            if (g_is_interrupted) break;
             console::set_display(console::reset);
             line = string_strip(line);
             if (line.empty()) {
@@ -323,10 +316,17 @@ int main(int argc, char ** argv) {
                 continue;
             }
             g_is_generating = true;
-            if (line.find("/image") == 0) {
+            if (line == "/image" || line.find("/image ") == 0) {
+                if (line.size() < 8) {
+                    LOG_ERR("ERR: Missing image filename\n");
+                    continue;
+                }
                 std::string image = line.substr(7);
-                images_fname.push_back(string_strip(image));
-                content += "<__image__>";
+                if (ctx.load_image(image)) {
+                    LOG("Image %s loaded\n", image.c_str());
+                    content += "<__image__>";
+                }
+                // else, error is already printed by libmtmd
                 continue;
             } else {
                 content += line;
@@ -334,24 +334,20 @@ int main(int argc, char ** argv) {
             common_chat_msg msg;
             msg.role = "user";
             msg.content = content;
-            int ret = eval_message(ctx, msg, images_fname, is_first_msg);
-            if (ret == 2) {
-                // non-fatal error
-                images_fname.clear();
-                content.clear();
-                continue;
-            }
+            int ret = eval_message(ctx, msg, is_first_msg);
             if (ret) {
                 return 1;
             }
+            if (g_is_interrupted) break;
             if (generate_response(ctx, smpl, n_predict)) {
                 return 1;
             }
-            images_fname.clear();
             content.clear();
             is_first_msg = false;
         }
     }
+    if (g_is_interrupted) LOG("\nInterrupted by user\n");
+    LOG("\n\n");
     llama_perf_context_print(ctx.lctx);
-    return 0;
+    return g_is_interrupted ? 130 : 0;
 }

examples/llava/mtmd.cpp

@@ -40,11 +40,14 @@ struct mtmd_context {
     llama_token tok_sli_img_end   = LLAMA_TOKEN_NULL; // single slice
     llama_token tok_row_end       = LLAMA_TOKEN_NULL; // end of row
 
+    bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
+
     // TODO @ngxson : add timings
 
     mtmd_context(const char * mmproj_fname,
                    const llama_model * text_model,
                    const mtmd_context_params & ctx_params) :
+        text_model   (text_model),
         print_timings(ctx_params.print_timings),
         n_threads    (ctx_params.n_threads),
         image_marker (ctx_params.image_marker)
@@ -56,9 +59,8 @@ struct mtmd_context {
         if (!ctx_clip) {
             throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
         }
-        this->text_model = text_model;
 
-        GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
+        use_mrope = clip_is_qwen2vl(ctx_clip);
 
         int minicpmv_version = clip_is_minicpmv(ctx_clip);
         if (minicpmv_version == 2) {
@@ -126,6 +128,7 @@ struct mtmd_image_tokens_data {
 struct mtmd_image_tokens {
     uint32_t nx; // number of tokens in x direction
     uint32_t ny; // number of tokens in y direction
+    bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position)
     uint32_t n_tokens() const { return nx * ny; }
     clip_image_f32_batch batch_f32; // preprocessed image patches
     std::string id; // optional user-defined ID, useful for KV cache tracking
@@ -176,6 +179,8 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
 
     std::string prompt_modified(text.text);
     std::string marker_modified(ctx->image_marker);
+    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
+
     // a bit hacky here, but works for now
     // for some models, we need to add prefix and suffix to the image embeddings
     if (clip_is_gemma3(ctx->ctx_clip)) {
@@ -183,12 +188,31 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
         // <start_of_image> ... (image embeddings) ... <end_of_image>
         marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
         string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    } else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+        // <|begin_of_image|> ... (image embeddings) ... <|end_of_image|>
+        marker_modified = "<|begin_of_image|>" + ctx->image_marker + "<|end_of_image|>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
+        // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
+        marker_modified = "<fake_token_around_image><global-img>" + ctx->image_marker + "<fake_token_around_image>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    } else if (proj_type == PROJECTOR_TYPE_PIXTRAL) {
+        // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
+        marker_modified = ctx->image_marker + "[IMG_END]";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+    }
+
+    else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
+        // <|vision_start|> ... (image embeddings) ... <|vision_end|>
+        marker_modified = "<|vision_start|>" + ctx->image_marker + "<|vision_end|>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
     }
 
     // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
-    // for glm-edge, we don't need to add because the tokens are already in the returned embeddings
-
-    // TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens
 
     std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
     output.clear();
@@ -212,7 +236,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
 
         for (auto & entry : batch_f32.entries) {
             mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-            image_tokens->nx = clip_n_patches(ctx->ctx_clip);
+            image_tokens->nx = clip_n_output_tokens(ctx->ctx_clip, entry.get());
             image_tokens->ny = 1;
             image_tokens->batch_f32.entries.push_back(std::move(entry));
             image_tokens->id = id;
@@ -229,7 +253,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
     };
 
     for (const auto & part : parts) {
-        //printf("tokenizing part: %s\n", part.c_str());
+        // printf("tokenizing part: %s\n", part.c_str());
         bool add_bos = &parts.front() == &part;
         auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special);
         if (tokens.empty()) {
@@ -306,9 +330,22 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                 }
 
             } else {
+                size_t n_tokens = 0;
+                for (const auto & entry : batch_f32.entries) {
+                    n_tokens += clip_n_output_tokens(ctx->ctx_clip, entry.get());
+                }
+
                 mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-                image_tokens->nx = clip_n_patches(ctx->ctx_clip) * batch_f32.entries.size(); // TODO @ngxson : use clip_n_patches_by_image
-                image_tokens->ny = 1; // TODO
+                if (ctx->use_mrope) {
+                    // for Qwen2VL, we need this information for M-RoPE decoding positions
+                    image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_clip, batch_f32.entries[0].get());
+                    image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_clip, batch_f32.entries[0].get());
+                    image_tokens->use_mrope_pos = true;
+                } else {
+                    // other models, we only need the total number of tokens
+                    image_tokens->nx = n_tokens;
+                    image_tokens->ny = 1;
+                }
                 image_tokens->batch_f32 = std::move(batch_f32);
                 image_tokens->id = bitmaps[i_img].id; // optional
 
@@ -316,11 +353,6 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                 LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
                 LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
 
-                if (clip_is_glm(ctx->ctx_clip)) {
-                    // glm-edge
-                    image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings
-                }
-
                 mtmd_input_chunk chunk{
                     MTMD_INPUT_CHUNK_TYPE_IMAGE,
                     {},
@@ -358,6 +390,13 @@ std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
     return image_tokens->id;
 }
 
+llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
+    if (image_tokens->use_mrope_pos) {
+        return 1; // for M-RoPE, the whole image is 1 in temporal dimension
+    }
+    return image_tokens->n_tokens();
+}
+
 int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
     int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
     ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
@@ -375,7 +414,7 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
         // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
         const auto & entries = image_tokens->batch_f32.entries;
         for (size_t i = 0; i < entries.size(); i++) {
-            int n_tokens_per_image = clip_n_patches(ctx->ctx_clip);
+            int n_tokens_per_image = clip_n_output_tokens(ctx->ctx_clip, entries[i].get());
             ok = clip_image_encode(
                 ctx->ctx_clip,
                 ctx->n_threads,
@@ -403,7 +442,7 @@ size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
         if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
             n_tokens += chunk.tokens_text.size();
         } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            n_tokens += chunk.tokens_image->n_tokens();
+            n_tokens += mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
         } else {
             GGML_ASSERT(false && "chunk type not supported");
         }
@@ -411,22 +450,38 @@ size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
     return n_tokens;
 }
 
+llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks) {
+    llama_pos n_pos = 0;
+    for (auto & chunk : chunks) {
+        if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            n_pos += chunk.tokens_text.size();
+        } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            n_pos += mtmd_image_tokens_get_n_pos(chunk.tokens_image.get());
+        } else {
+            GGML_ASSERT(false && "chunk type not supported");
+        }
+    }
+    return n_pos;
+}
+
 // helper struct to make working with embd batch easier
 // note: this will be removed after llama_batch_ext refactoring
 struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
     std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
     std::vector<int32_t>        n_seq_id;
     std::vector<llama_seq_id>   seq_id_0;
     std::vector<llama_seq_id *> seq_ids;
     std::vector<int8_t>         logits;
     llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
-        pos     .resize(n_tokens);
+    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        pos     .resize(n_tokens * n_pos_per_embd);
         n_seq_id.resize(n_tokens);
         seq_ids .resize(n_tokens + 1);
         logits  .resize(n_tokens);
         seq_id_0.resize(1);
-        seq_id_0[0] = seq_id;
         seq_ids [n_tokens] = nullptr;
         batch = {
             /*n_tokens       =*/ n_tokens,
@@ -437,13 +492,64 @@ struct decode_embd_batch {
             /*seq_id         =*/ seq_ids.data(),
             /*logits         =*/ logits.data(),
         };
-        for (int i = 0; i < n_tokens; i++) {
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
             batch.pos     [i] = pos_0 + i;
             batch.n_seq_id[i] = 1;
             batch.seq_id  [i] = seq_id_0.data();
             batch.logits  [i] = false;
         }
     }
+
+    void set_position_mrope(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int y = 0; y < ny; y++) {
+            for (int x = 0; x < nx; x++) {
+                int i = y * nx + x;
+                pos[i                     ] = pos_0;
+                pos[i + batch.n_tokens    ] = pos_0 + y;
+                pos[i + batch.n_tokens * 2] = pos_0 + x;
+                pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
+            }
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.resize(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                auto src = pos.begin() + i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(), src, src + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
 };
 
 int32_t mtmd_helper_eval(mtmd_context * ctx,
@@ -456,6 +562,7 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
     llama_pos n_past = pos0;
     llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
     int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
+    int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
 
     for (auto & chunk : chunks) {
         bool is_last = &chunk == &chunks.back();
@@ -483,7 +590,7 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             }
 
         } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
-            GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
+            GGML_ASSERT(!is_last && "logits for last image chunk is not yet supported");
             GGML_ASSERT(chunk.tokens_image != nullptr);
             int64_t t0 = ggml_time_ms();
             if (ctx->print_timings) {
@@ -503,6 +610,16 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             int32_t i_batch = 0;
             int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
             float * embd = mtmd_get_output_embd(ctx);
+            decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+
+            const int nx = mtmd_image_tokens_get_nx(chunk.tokens_image.get());
+            const int ny = mtmd_image_tokens_get_ny(chunk.tokens_image.get());
+
+            if (mtmd_decode_use_mrope(ctx)) {
+                batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
+            } else {
+                batch_embd.set_position_normal(n_past, seq_id);
+            }
 
             if (mtmd_decode_use_non_causal(ctx)) {
                 llama_set_causal_attn(lctx, false);
@@ -510,15 +627,14 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
             }
 
             while (i_batch < n_img_batches) { // split into batches
-                int32_t pos_offset = i_batch*n_batch;
-                int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
-                float * embd_batch = embd + pos_offset*n_mmproj_embd;
-                decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
+                int pos_offset = i_batch*n_batch;
+                int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
+                llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
 
-                printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
+                LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
 
                 int64_t t1 = ggml_time_ms();
-                ret = llama_decode(lctx, batch_img.batch);
+                ret = llama_decode(lctx, batch_embd_view);
                 if (ret != 0) {
                     LOG_ERR("failed to decode image\n");
                     llama_set_causal_attn(lctx, true); // restore causal attn
@@ -531,9 +647,11 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
                 }
 
                 i_batch++;
-                n_past += n_tokens_batch;
             }
 
+            // for mrope, one image is one single **temporal** position
+            n_past += mtmd_decode_use_mrope(ctx) ? 1 : n_tokens;
+
             if (mtmd_decode_use_non_causal(ctx)) {
                 llama_set_causal_attn(lctx, true);
             }
@@ -581,6 +699,10 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
     return false;
 }
 
+bool mtmd_decode_use_mrope(mtmd_context * ctx) {
+    return ctx->use_mrope;
+}
+
 void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
     mtmd_image_tokens_free(val);
 }

examples/llava/mtmd.h

@@ -102,6 +102,7 @@ MTMD_API size_t      mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * im
 MTMD_API size_t      mtmd_image_tokens_get_nx(const mtmd_image_tokens * image_tokens);
 MTMD_API size_t      mtmd_image_tokens_get_ny(const mtmd_image_tokens * image_tokens);
 MTMD_API std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
+MTMD_API llama_pos   mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens); // number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
 MTMD_API void        mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
 
 // returns 0 on success
@@ -114,15 +115,21 @@ MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 // whether we need to set non-causal mask before llama_decode
 MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
 
+// whether the current model use M-RoPE for llama_decode
+MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);
+
 
 
 //
 // helper functions (can be implemented based on other functions)
 //
 
-// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
+// helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache
 MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks);
 
+// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
+MTMD_API llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks);
+
 // helper function that automatically:
 // 1. run llama_decode() on text chunks
 // 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()

examples/llava/qwen2_vl_surgery.py

@@ -1,14 +1,16 @@
 import argparse
-from typing import Dict
+from typing import Dict, List, Optional
 
 import torch
 import numpy as np
 from gguf import *
 from transformers import (
-    Qwen2VLForConditionalGeneration,
-    Qwen2VLProcessor,
     AutoProcessor,
-    Qwen2VLConfig
+    Qwen2VLConfig,
+    Qwen2VLProcessor,
+    Qwen2VLForConditionalGeneration,
+    Qwen2_5_VLConfig, # type: ignore[reportAttributeAccessIssue]
+    Qwen2_5_VLForConditionalGeneration, # type: ignore[reportAttributeAccessIssue]
 )
 
 
@@ -19,61 +21,93 @@ def k(raw_key: str, arch: str) -> str:
     return raw_key.format(arch=arch)
 
 
-def to_gguf_name(name: str) -> str:
-    og = name
-    name = name.replace("text_model", "t").replace("vision_model", "v")
-    name = name.replace("blocks", "blk").replace("embeddings.", "")
-    name = name.replace("attn.", "attn_")
-    name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
-    # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
-    name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-    name = name.replace("merger.mlp", 'mm')
-    print(f"[to_gguf_name] {og} --> {name}")
-    return name
+def get_n_wa_pattern(fullatt_block_indexes: Optional[List[int]]):
+    if fullatt_block_indexes is None:
+        return 0
+    n_wa = fullatt_block_indexes[0]
+    for a, b in zip(fullatt_block_indexes, fullatt_block_indexes[1:]):
+        if b - a - 1 != n_wa:
+            raise ValueError(
+                f"window/full attention layer should have fix pattern of "
+                f"for each full-attention layer followed by {n_wa} window-attention layers"
+            )
+    return n_wa + 1
 
 
-def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
-    vision_model = qwen2vl.visual
-    tensor_map = {}
-    for name, ten in vision_model.state_dict().items():
-        ten = ten.numpy()
-        if 'qkv' in name:
-            if ten.ndim == 2: # weight
-                c3, _ = ten.shape
-            else:             # bias
-                c3 = ten.shape[0]
-            assert c3 % 3 == 0
-            c = c3 // 3
-            wq = ten[:c]
-            wk = ten[c: c * 2]
-            wv = ten[c * 2:]
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
-        elif 'merger' in name:
-            if name.endswith("ln_q.weight"):
-                tensor_map['v.post_ln.weight'] = ten
-            elif name.endswith("ln_q.bias"):
-                tensor_map['v.post_ln.bias'] = ten
+class VL2:
+
+    @staticmethod
+    def to_gguf_name(name: str) -> str:
+        og = name
+        name = name.replace("text_model", "t").replace("vision_model", "v")
+        name = name.replace("blocks", "blk").replace("embeddings.", "")
+        name = name.replace("attn.", "attn_")
+        name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
+        # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
+        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+        name = name.replace("merger.mlp", 'mm')
+        print(f"[to_gguf_name] {og} --> {name}")
+        return name
+
+    @classmethod
+    def find_vision_tensors(cls, qwen2vl, dtype) -> Dict[str, np.ndarray]:
+        vision_model = qwen2vl.visual
+        tensor_map = {}
+        for name, ten in vision_model.state_dict().items():
+            ten = ten.numpy()
+            if 'qkv' in name:
+                if ten.ndim == 2: # weight
+                    c3, _ = ten.shape
+                else:             # bias
+                    c3 = ten.shape[0]
+                assert c3 % 3 == 0
+                c = c3 // 3
+                wq = ten[:c]
+                wk = ten[c: c * 2]
+                wv = ten[c * 2:]
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
+            elif 'merger' in name:
+                if name.endswith("ln_q.weight"):
+                    tensor_map['v.post_ln.weight'] = ten
+                elif name.endswith("ln_q.bias"):
+                    tensor_map['v.post_ln.bias'] = ten
+                else:
+                    # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
+                    tensor_map[cls.to_gguf_name(name)] = ten
+            elif 'patch_embed.proj.weight' in name:
+                # NOTE: split Conv3D into Conv2Ds
+                c1, c2, kt, kh, kw = ten.shape
+                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+                tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
+                tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
             else:
-                # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
-                tensor_map[to_gguf_name(name)] = ten
-        elif 'patch_embed.proj.weight' in name:
-            # NOTE: split Conv3D into Conv2Ds
-            c1, c2, kt, kh, kw = ten.shape
-            assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-            tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
-            tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
-        else:
-            tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}")] = ten
 
-    for new_name, ten in tensor_map.items():
-        if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
-            tensor_map[new_name] = ten.astype(np.float32)
-        else:
-            tensor_map[new_name] = ten.astype(dtype)
-    tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
-    return tensor_map
+        for new_name, ten in tensor_map.items():
+            if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
+                tensor_map[new_name] = ten.astype(np.float32)
+            else:
+                tensor_map[new_name] = ten.astype(dtype)
+        tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
+        return tensor_map
+
+
+class VL25(VL2):
+
+    @staticmethod
+    def to_gguf_name(name: str) -> str:
+        og = name
+        name = name.replace("text_model", "t").replace("vision_model", "v")
+        name = name.replace("blocks", "blk").replace("embeddings.", "")
+        name = name.replace("attn.", "attn_")
+        name = name.replace("mlp.down_proj", "ffn_down").replace("mlp.up_proj", "ffn_up")
+        name = name.replace("mlp.gate_proj", "ffn_gate").replace("proj.", "out.")
+        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+        name = name.replace("merger.mlp", 'mm')
+        print(f"[vl25][to_gguf_name] {og} --> {name}")
+        return name
 
 
 def main(args):
@@ -82,7 +116,7 @@ def main(args):
         np_dtype = np.float32
         ftype = 0
     elif args.data_type == 'fp16':
-        dtype = torch.float32
+        dtype = torch.float16
         np_dtype = np.float16
         ftype = 1
     else:
@@ -92,11 +126,18 @@ def main(args):
     model_path = ""
     model_name = args.model_name
     print("model_name: ", model_name)
-    qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
-        model_name, torch_dtype=dtype, device_map="cpu"
-    )
-    cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-    vcfg = cfg.vision_config
+    if args.model_type == "qwen2vl":
+        qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=dtype, device_map="cpu"
+        )
+        cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+        vcfg = cfg.vision_config
+    else:
+        qwen2vl = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=dtype, device_map="cpu"
+        )
+        cfg: Qwen2_5_VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+        vcfg = cfg.vision_config
 
     if os.path.isdir(model_name):
         local_model = True
@@ -113,7 +154,6 @@ def main(args):
     fout.add_bool("clip.has_text_encoder", False)
     fout.add_bool("clip.has_vision_encoder", True)
     fout.add_bool("clip.has_qwen2vl_merger", True)
-    fout.add_string("clip.projector_type", "qwen2vl_merger")
 
     print(cfg.vision_config)
     if 'silu' in cfg.vision_config.hidden_act.lower():
@@ -125,14 +165,25 @@ def main(args):
     else:
         raise ValueError()
 
-    tensor_map = find_vision_tensors(qwen2vl, np_dtype)
+    if args.model_type == "qwen2.5vl":
+        fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
+        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
+        fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)
+        fout.add_string("clip.projector_type", "qwen2.5vl_merger")
+    else:
+        fout.add_string("clip.projector_type", "qwen2vl_merger")
+        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
+        fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
+
+    if args.model_type == "qwen2.5vl":
+        tensor_map = VL25.find_vision_tensors(qwen2vl, np_dtype)
+    else:
+        tensor_map = VL2.find_vision_tensors(qwen2vl, np_dtype)
     for name, data in tensor_map.items():
         fout.add_tensor(name, data)
 
     fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
     fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
-    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
-    fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
     fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
@@ -160,6 +211,7 @@ def main(args):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
+    parser.add_argument("--model_type", nargs='?', choices=['qwen2vl', 'qwen2.5vl'], default="qwen2vl")
     parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
     args = parser.parse_args()
     main(args)

examples/llava/qwen2vl-test.cpp

@@ -23,7 +23,12 @@
 #include <algorithm>
 #include <iostream>
 #include <fstream>
+#include <limits>
+#include <cassert>
+#include <cmath>
 
+// THIS FILE IS ONLY USED FOR TESTING THE QWEN2VL MODEL
+// IT IS NOT A PRODUCTION CODE
 
 static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
                                      int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {
@@ -89,20 +94,12 @@ static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct lla
 
 static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past, int * st_pos_id) {
     int N = (int) tokens.size();
-    std::vector<llama_pos> pos;
     for (int i = 0; i < N; i += n_batch) {
         int n_eval = (int) tokens.size() - i;
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
         auto batch = llama_batch_get_one(&tokens[i], n_eval);
-        // TODO: add mrope pos ids somewhere else
-        pos.resize(batch.n_tokens * 4);
-        std::fill(pos.begin(), pos.end(), 0);
-        for (int j = 0; j < batch.n_tokens * 3; j ++) {
-            pos[j] = *st_pos_id + (j % batch.n_tokens);
-        }
-        batch.pos = pos.data();
 
         if (llama_decode(ctx_llama, batch)) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
@@ -367,14 +364,14 @@ static void debug_test_mrope_2d() {
     // 1. Initialize backend
     ggml_backend_t backend = NULL;
     std::string backend_name = "";
-#ifdef GGML_USE_CUDA
-    fprintf(stderr, "%s: using CUDA backend\n", __func__);
-    backend = ggml_backend_cuda_init(0); // init device 0
-    backend_name = "cuda";
-    if (!backend) {
-        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
-    }
-#endif
+// #ifdef GGML_USE_CUDA
+//     fprintf(stderr, "%s: using CUDA backend\n", __func__);
+//     backend = ggml_backend_cuda_init(0); // init device 0
+//     backend_name = "cuda";
+//     if (!backend) {
+//         fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+//     }
+// #endif
     // if there aren't GPU Backends fallback to CPU backend
     if (!backend) {
         backend = ggml_backend_cpu_init();
@@ -483,28 +480,82 @@ static void debug_test_mrope_2d() {
     ggml_backend_free(backend);
 }
 
-static void debug_dump_img_embed(struct llava_context * ctx_llava) {
-    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
-    int ne = n_embd * 4;
-    float vals[56 * 56 * 3];
+enum model_output_type {
+    conv3d,
+    patch_embed,
+    patch_win_attn_scatter,
+    first_attn_layer,
+    last_attn_layer,
+    attn_softmax,
+    final_layer,
+};
+
+static void debug_dump_img_embed(struct llava_context * ctx_llava, model_output_type output_type) {
+    constexpr int ih = 140;
+    constexpr int iw = 196;
+    // constexpr int ih = 56;
+    // constexpr int iw = 56;
+    // int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
+    int n_embd  = 1280;
+    int merge = 1;
+    if (output_type == model_output_type::final_layer) {
+        n_embd  = 2048;
+        merge = 2;
+    }
+    else if (output_type == model_output_type::attn_softmax) {
+        merge = 1;
+        n_embd = (ih/14/merge) * (iw/14/merge) * 16;
+    }
+
+    int ne = (ih/14/merge) * (iw/14/merge) * n_embd;
+    float vals[iw * ih * 3];
     // float embd[ne];
     std::vector<float> embd;
     embd.resize(ne);
 
-    for (int i = 0; i < 56*56; i++)
+    for (int i = 0; i < iw*ih; i++)
     {
         for (int c = 0; c < 3; c++)
-            vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56);
+            vals[i * 3 + c] = (float)i / (iw*ih);
     }
 
-    clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data());
+    clip_encode_float_image(ctx_llava->ctx_clip, 8, vals, ih, iw, embd.data());
 
-    std::ofstream outFile("img_embed.bin", std::ios::binary);
+    std::string file_postfix = "";
+    switch (output_type)
+    {
+    case model_output_type::conv3d:
+        file_postfix = "conv3d";
+        break;
+    case model_output_type::patch_embed:
+        file_postfix = "patch_embed";
+        break;
+    case model_output_type::patch_win_attn_scatter:
+        file_postfix = "scatter";
+        break;
+    case model_output_type::first_attn_layer:
+        file_postfix = "first_attn";
+        break;
+    case model_output_type::last_attn_layer:
+        file_postfix = "last_attn";
+        break;
+    case model_output_type::attn_softmax:
+        file_postfix = "attn_softmax";
+        break;
+    case model_output_type::final_layer:
+        file_postfix = "final";
+        break;
+    default:
+        break;
+    }
+    auto output_path = "img_embed_" + file_postfix + ".bin";
+
+    std::ofstream outFile(output_path, std::ios::binary);
     if (outFile.is_open()) {
         outFile.write(reinterpret_cast<const char*>(embd.data()), ne * sizeof(float));
 
         outFile.close();
-        std::cout << "Data successfully written to mrope.bin" << std::endl;
+        std::cout << "Data successfully written to ::[ " << output_path << std::endl;
     } else {
         std::cerr << "Error opening file!" << std::endl;
     }
@@ -551,8 +602,9 @@ int main(int argc, char ** argv) {
     } else if (params.image[0].empty()) {
         auto ctx_llava = llava_init_context(&params, model);
 
-        debug_test_mrope_2d();
-        debug_dump_img_embed(ctx_llava);
+        // debug_test_mrope_2d();
+        debug_dump_img_embed(ctx_llava, model_output_type::final_layer);
+        // debug_dump_img_embed(ctx_llava, model_output_type::last_attn_layer);
 
         llama_perf_context_print(ctx_llava->ctx_llama);
         ctx_llava->model = NULL;

examples/llava/tests.sh

@@ -13,6 +13,14 @@ mkdir -p $SCRIPT_DIR/output
 PROJ_ROOT="$SCRIPT_DIR/../.."
 cd $PROJ_ROOT
 
+# Check if the first argument is "big", then run test with big models
+# This is useful if we're running the script on a larger machine, so we can test the big models
+RUN_BIG_TESTS=false
+if [ "${1:-}" = "big" ]; then
+    RUN_BIG_TESTS=true
+    echo "Include BIG models..."
+fi
+
 ###############
 
 arr_bin=()
@@ -28,6 +36,15 @@ add_test() {
     arr_tmpl+=("$tmpl")
 }
 
+add_test_big() {
+    if [ "$RUN_BIG_TESTS" = true ]; then
+        add_test "$@"
+    fi
+}
+
+add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
+add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
 add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"               "deepseek"
 add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
@@ -37,9 +54,20 @@ add_test "llama-mtmd-cli"  "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # model from openbmb is corrupted
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
-add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
 
-# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"  # this model has broken chat template, not usable
+# to test the big models, run: ./tests.sh big
+add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+add_test_big "llama-mtmd-cli" "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" "mistral-v7"
+
+# these models always give the wrong answer, not sure why
+# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"
+# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-256M-Instruct-GGUF:Q8_0"
+# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF:Q8_0"
+
+# this model has broken chat template, not usable
+# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
 
 ###############
 

examples/quantize-stats/CMakeLists.txt

@@ -1,6 +0,0 @@
-set(TARGET llama-quantize-stats)
-add_executable(${TARGET} quantize-stats.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/rpc/rpc-server.cpp

@@ -22,6 +22,7 @@
 
 #include "ggml-rpc.h"
 #ifdef _WIN32
+#  define NOMINMAX
 #  define DIRECTORY_SEPARATOR '\\'
 #  include <locale>
 #  include <windows.h>
@@ -37,6 +38,8 @@
 #include <stdio.h>
 #include <vector>
 #include <filesystem>
+#include <algorithm>
+#include <thread>
 
 namespace fs = std::filesystem;
 
@@ -150,12 +153,14 @@ struct rpc_server_params {
     int         port        = 50052;
     size_t      backend_mem = 0;
     bool        use_cache   = false;
+    int         n_threads   = std::max(1U, std::thread::hardware_concurrency()/2);
 };
 
 static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
     fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help                show this help message and exit\n");
+    fprintf(stderr, "  -t,      --threads        number of threads for the CPU backend (default: %d)\n", params.n_threads);
     fprintf(stderr, "  -H HOST, --host HOST      host to bind to (default: %s)\n", params.host.c_str());
     fprintf(stderr, "  -p PORT, --port PORT      port to bind to (default: %d)\n", params.port);
     fprintf(stderr, "  -m MEM,  --mem MEM        backend memory size (in MB)\n");
@@ -172,6 +177,15 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
                 return false;
             }
             params.host = argv[i];
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
+                return false;
+            }
         } else if (arg == "-p" || arg == "--port") {
             if (++i >= argc) {
                 return false;
@@ -199,7 +213,7 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
     return true;
 }
 
-static ggml_backend_t create_backend() {
+static ggml_backend_t create_backend(const rpc_server_params & params) {
     ggml_backend_t backend = NULL;
 #ifdef GGML_USE_CUDA
     fprintf(stderr, "%s: using CUDA backend\n", __func__);
@@ -231,6 +245,7 @@ static ggml_backend_t create_backend() {
     if (!backend) {
         fprintf(stderr, "%s: using CPU backend\n", __func__);
         backend = ggml_backend_cpu_init();
+        ggml_backend_cpu_set_n_threads(backend, params.n_threads);
     }
     return backend;
 }
@@ -275,7 +290,7 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "\n");
     }
 
-    ggml_backend_t backend = create_backend();
+    ggml_backend_t backend = create_backend(params);
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");
         return 1;
@@ -289,8 +304,9 @@ int main(int argc, char * argv[]) {
         get_backend_memory(&free_mem, &total_mem);
     }
     const char * cache_dir = nullptr;
-    std::string cache_dir_str = fs_get_cache_directory() + "rpc/";
+    std::string cache_dir_str;
     if (params.use_cache) {
+        cache_dir_str = fs_get_cache_directory() + "rpc/";
         if (!fs_create_directory_with_parents(cache_dir_str)) {
             fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str());
             return 1;

examples/server/README.md

@@ -154,7 +154,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
-| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
+| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
 | `--slots` | enable slots monitoring endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |

examples/server/public_legacy/json-schema-to-grammar.mjs

@@ -2,6 +2,9 @@
 const SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}';
 
 function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
+  if (maxItems == 0) {
+    return '';
+  }
   if (minItems === 0 && maxItems === 1) {
     return `${itemRule}?`;
   }

examples/server/utils.hpp

@@ -642,9 +642,31 @@ static json oaicompat_completion_params_parse(
         throw std::runtime_error("Cannot use custom grammar constraints with tools.");
     }
 
+    // if the assistant message appears at the end of list, we do not add end-of-turn token
+    // for ex. this can be useful to modify the reasoning process in reasoning models
+    bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant";
+    common_chat_msg last_message;
+    if (prefill_assistant_message) {
+        last_message = inputs.messages.back();
+        inputs.messages.pop_back();
+
+        /* sanity check, max one assistant message at the end of the list */
+        if (!inputs.messages.empty() && inputs.messages.back().role == "assistant"){
+            throw std::runtime_error("Cannot have 2 or more assistant messages at the end of the list.");
+        }
+
+        inputs.extract_reasoning = false;
+        inputs.add_generation_prompt = true;
+    }
+
     // Apply chat template to the list of messages
     auto chat_params = common_chat_templates_apply(tmpls, inputs);
 
+    /* Append assistant prefilled message */
+    if (prefill_assistant_message) {
+         chat_params.prompt += last_message.content;
+    }
+
     llama_params["chat_format"]      = static_cast<int>(chat_params.format);
     llama_params["prompt"]           = chat_params.prompt;
     if (!chat_params.grammar.empty()) {

ggml/CMakeLists.txt

@@ -360,3 +360,27 @@ write_basic_package_version_file(
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
+
+if (MSVC)
+    set(MSVC_WARNING_FLAGS
+        /wd4005  # Macro redefinition
+        /wd4244  # Conversion from one type to another type, possible loss of data
+        /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+    )
+    function(disable_msvc_warnings target_name)
+        if(TARGET ${target_name})
+            target_compile_options(${target_name} PRIVATE ${MSVC_WARNING_FLAGS})
+        endif()
+    endfunction()
+
+    disable_msvc_warnings(ggml-base)
+    disable_msvc_warnings(ggml)
+    disable_msvc_warnings(ggml-cpu)
+    disable_msvc_warnings(ggml-cpu-x64)
+    disable_msvc_warnings(ggml-cpu-sse42)
+    disable_msvc_warnings(ggml-cpu-sandybridge)
+    disable_msvc_warnings(ggml-cpu-haswell)
+    disable_msvc_warnings(ggml-cpu-skylakex)
+    disable_msvc_warnings(ggml-cpu-icelake)
+    disable_msvc_warnings(ggml-cpu-alderlake)
+endif()

ggml/include/ggml-cpp.h

@@ -24,7 +24,7 @@ typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
 
 struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
 
-typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+typedef std::unique_ptr<ggml_gallocr, ggml_gallocr_deleter> ggml_gallocr_ptr;
 
 // ggml-backend
 

ggml/include/ggml-cpu.h

@@ -133,6 +133,11 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_bf16_to_fp32(const ggml_bf16_t *, float *, int64_t);
+
 #ifdef __cplusplus
 }
 #endif

ggml/include/ggml-rpc.h

@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    2
 #define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16

ggml/include/ggml.h

@@ -393,8 +393,8 @@ extern "C" {
 
     // precision
     enum ggml_prec {
-        GGML_PREC_DEFAULT,
-        GGML_PREC_F32,
+        GGML_PREC_DEFAULT =  0, // stored as ggml_tensor.op_params, 0 by default
+        GGML_PREC_F32     = 10,
     };
 
     // model file types
@@ -481,6 +481,7 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
@@ -677,6 +678,9 @@ extern "C" {
     GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
     GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
 
+    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
+    GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -1660,7 +1664,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // depthwise
+    // depthwise (via im2col and mul_mat)
     GGML_API struct ggml_tensor * ggml_conv_2d_dw(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,  // convolution kernel
@@ -1672,6 +1676,22 @@ extern "C" {
             int                  d0,  // dilation dimension 0
             int                  d1); // dilation dimension 1
 
+    // Depthwise 2D convolution
+    // may be faster than ggml_conv_2d_dw, but not available in all backends
+    // a:   KW    KH    1    C    convolution kernel
+    // b:   W     H     C    N    input data
+    // res: W_out H_out C    N
+    GGML_API struct ggml_tensor * ggml_conv_2d_dw_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/ggml-alloc.c

@@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
     size_t node_size = 0;
     if (!node->data && !node->view_src) {
-        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        // If we previously had data but don't now then reallocate
+        if (talloc->buffer_id < 0) {
+            return false;
+        }
         node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
     }
     return talloc->size_max >= node_size;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -352,10 +352,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
             message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
+            list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
             message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+            list(APPEND ARCH_FLAGS -march=z16)
+        elseif (${S390X_M} MATCHES "9175|9176")
+            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            message(STATUS "z17 target")
+            list(APPEND ARCH_FLAGS -march=z17)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")

ggml/src/ggml-cpu/ggml-cpu.c

@@ -215,7 +215,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
     },
     [GGML_TYPE_F16] = {
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp16,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
         .vec_dot_type             = GGML_TYPE_F16,
         .nrows                    = 1,
@@ -356,7 +356,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q8_K,
     },
     [GGML_TYPE_BF16] = {
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_bf16,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
@@ -1932,6 +1932,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
+        case GGML_OP_CONV_2D_DW:
+            {
+                ggml_compute_forward_conv_2d_dw(params, tensor);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
                 ggml_compute_forward_conv_transpose_2d(params, tensor);
@@ -2268,6 +2272,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
@@ -3161,6 +3166,93 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        __m512 x_vec = _mm512_loadu_ps(x + i);
+        __m256i y_vec = _mm512_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm256_storeu_si256((__m256i *)(y + i), y_vec);
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        __m256 x_vec = _mm256_loadu_ps(x + i);
+        __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storeu_si128((__m128i *)(y + i), y_vec);
+    }
+    for (; i + 3 < n; i += 4) {
+        __m128 x_vec = _mm_loadu_ps(x + i);
+        __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storel_epi64((__m128i *)(y + i), y_vec);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(x[i]);
+    }
+}
+
+void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        __m256i x_vec = _mm256_loadu_si256((const __m256i *)(x + i));
+        __m512 y_vec = _mm512_cvtph_ps(x_vec);
+        _mm512_storeu_ps(y + i, y_vec);
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        __m128i x_vec = _mm_loadu_si128((const __m128i *)(x + i));
+        __m256 y_vec = _mm256_cvtph_ps(x_vec);
+        _mm256_storeu_ps(y + i, y_vec);
+    }
+    for (; i + 3 < n; i += 4) {
+        __m128i x_vec = _mm_loadl_epi64((const __m128i *)(x + i));
+        __m128 y_vec = _mm_cvtph_ps(x_vec);
+        _mm_storeu_ps(y + i, y_vec);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+}
+
+void ggml_cpu_fp32_to_bf16(const float * x, ggml_bf16_t * y, int64_t n) {
+    int64_t i = 0;
+    for (; i < n; ++i) {
+        y[i] = GGML_FP32_TO_BF16(x[i]);
+    }
+}
+
+void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__AVX2__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i,
+                        _mm512_castsi512_ps(
+                            _mm512_slli_epi32(
+                                _mm512_cvtepu16_epi32(
+                                    _mm256_loadu_si256(
+                                        (const __m256i *)(x + i))),
+                                16)));
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i,
+                        _mm256_castsi256_ps(
+                            _mm256_slli_epi32(
+                                _mm256_cvtepu16_epi32(
+                                    _mm_loadu_si128(
+                                        (const __m128i *)(x + i))),
+                                16)));
+    }
+#endif
+    for (; i < n; i++) {
+        y[i] = GGML_BF16_TO_FP32(x[i]);
+    }
+}
 
 int ggml_cpu_has_avx(void) {
 #if defined(__AVX__)

ggml/src/ggml-cpu/ops.cpp

@@ -4222,7 +4222,7 @@ static void ggml_compute_forward_get_rows_f16(
 
         GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
-        ggml_fp16_to_fp32_row(
+        ggml_cpu_fp16_to_fp32(
             (const ggml_fp16_t*) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
@@ -4263,7 +4263,7 @@ static void ggml_compute_forward_get_rows_bf16(
 
         GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
-        ggml_bf16_to_fp32_row(
+        ggml_cpu_bf16_to_fp32(
             (const ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
                         (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
@@ -6064,6 +6064,178 @@ void ggml_compute_forward_conv_transpose_2d(
     }
 }
 
+// ggml_compute_forward_conv_2d_dw
+
+struct ggml_conv_2d_dw_params {
+    int64_t channels;
+    int64_t batch;
+    int64_t src_w;
+    int64_t src_h;
+    int64_t dst_w;
+    int64_t dst_h;
+    int64_t knl_w;
+    int64_t knl_h;
+    int stride_x;
+    int stride_y;
+    int pad_x;
+    int pad_y;
+    int dilation_x;
+    int dilation_y;
+};
+
+static void ggml_compute_forward_conv_2d_dw_cwhn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t c = p.channels;
+    const float * knl_data = (const float *)kernel->data;
+
+    const int64_t rows_total = p.dst_h * p.batch;
+    const int64_t rows_per_thread = (rows_total + params->nth - 1) / params->nth;
+    const int64_t row_start = params->ith * rows_per_thread;
+    const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
+
+#ifdef GGML_SIMD
+    const int64_t pkg_size = GGML_F32_EPR;
+    const int64_t pkg_count = c / pkg_size;
+    const int64_t c_pkg_end = pkg_count * pkg_size;
+#else
+    const int64_t c_pkg_end = 0;
+#endif
+
+    for (int64_t row = row_start; row < row_end; ++row) {
+        const int64_t dst_y = row % p.dst_h;
+        const float * src_data = (const float *)src->data + (row / p.dst_h) * p.src_w * p.src_h * c;
+        for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+            float * dst_data = (float *)dst->data + (row * p.dst_w + dst_x) * c;
+            const int64_t src_y_base = dst_y * p.stride_y - p.pad_y;
+            const int64_t src_x_base = dst_x * p.stride_x - p.pad_x;
+
+#ifdef GGML_SIMD
+            // Vectorized loop
+            for (int64_t c_i = 0; c_i < c_pkg_end; c_i += pkg_size) {
+                GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        GGML_F32_VEC k = GGML_F32_VEC_LOAD(knl_data + (knl_y * p.knl_w + knl_x) * c + c_i);
+                        GGML_F32_VEC s = GGML_F32_VEC_LOAD(src_data + (src_y * p.src_w + src_x) * c + c_i);
+                        sum = GGML_F32_VEC_FMA(sum, k, s);
+                    }
+                }
+                GGML_F32_VEC_STORE(dst_data + c_i, sum);
+            }
+#endif
+            // Scalar loop
+            for (int64_t c_i = c_pkg_end; c_i < c; ++c_i) {
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[(knl_y * p.knl_w + knl_x) * c + c_i]
+                             * src_data[(src_y * p.src_w + src_x) * c + c_i];
+                    }
+                }
+                dst_data[c_i] = sum;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_conv_2d_dw_whcn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t n = p.channels * p.batch;
+    const int64_t per_thread = (n + params->nth - 1) / params->nth;
+    const int64_t start = params->ith * per_thread;
+    const int64_t end = MIN(start + per_thread, n);
+
+    for (int64_t i = start; i < end; ++i) {
+        const float * knl_data = (const float *)kernel->data + (i % p.channels) * p.knl_w * p.knl_h;
+        const float * src_data = (const float *)src->data + i * p.src_w * p.src_h;
+        float * dst_data = (float *)dst->data + i * p.dst_w * p.dst_h;
+
+        for (int64_t dst_y = 0; dst_y < p.dst_h; ++dst_y) {
+            for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[knl_y * p.knl_w + knl_x]
+                             * src_data[src_y * p.src_w + src_x];
+                    }
+                }
+                dst_data[dst_y * p.dst_w + dst_x] = sum;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_2d_dw(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * src = dst->src[1];
+    ggml_conv_2d_dw_params p;
+    p.channels = src->ne[2];
+    p.batch = src->ne[3];
+    p.src_w = src->ne[0];
+    p.src_h = src->ne[1];
+    p.dst_w = dst->ne[0];
+    p.dst_h = dst->ne[1];
+    p.knl_w = kernel->ne[0];
+    p.knl_h = kernel->ne[1];
+    p.stride_x = dst->op_params[0];
+    p.stride_y = dst->op_params[1];
+    p.pad_x = dst->op_params[2];
+    p.pad_y = dst->op_params[3];
+    p.dilation_x = dst->op_params[4];
+    p.dilation_y = dst->op_params[5];
+
+    GGML_ASSERT(kernel->ne[3] == p.channels);
+    GGML_ASSERT(dst->ne[3] == p.batch);
+
+    if (ggml_is_contiguous(src)) {
+        ggml_compute_forward_conv_2d_dw_whcn(params, src, kernel, dst, p);
+    } else if (ggml_is_contiguous_channels(src)) {
+        // kernel should also have channels most contiguous in memory
+        GGML_ASSERT(kernel->nb[0] >= kernel->nb[2] && kernel->nb[1] >= kernel->nb[0]);
+        ggml_compute_forward_conv_2d_dw_cwhn(params, src, kernel, dst, p);
+    } else {
+        GGML_ABORT("non-contiguous memory layout not supported");
+    }
+}
+
 // ggml_compute_forward_pool_1d_sk_p0
 
 static void ggml_compute_forward_pool_1d_sk_p0(

ggml/src/ggml-cpu/ops.h

@@ -65,6 +65,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/simd-mappings.h

@@ -341,7 +341,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F32_EPR  4
 
 #define GGML_F32x4              vector float
-#define GGML_F32x4_ZERO         0.0f
+#define GGML_F32x4_ZERO         {0.0f}
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
 #define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)

ggml/src/ggml-cuda/CMakeLists.txt

@@ -133,6 +133,7 @@ if (CUDAToolkit_FOUND)
                 COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
                 OUTPUT_VARIABLE CUDA_CCVER
                 ERROR_QUIET
+                OUTPUT_STRIP_TRAILING_WHITESPACE
             )
         else()
             if (CUDA_CCFULLVER MATCHES Apple)
@@ -143,7 +144,7 @@ if (CUDAToolkit_FOUND)
             string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
         endif()
 
-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+        message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
         list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later

ggml/src/ggml-cuda/common.cuh

@@ -78,13 +78,13 @@
 // Moore Threads
 #define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
 
-#define GGML_CUDA_CC_QY1  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
-#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
 #ifdef __CUDA_ARCH_LIST__

ggml/src/ggml-cuda/convert.cu

@@ -1,6 +1,8 @@
 #include "convert.cuh"
 #include "dequantize.cuh"
 
+#include <cstdint>
+
 #define CUDA_Q8_0_NE_ALIGN 2048
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
@@ -570,30 +572,46 @@ static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t
 }
 
 template <typename src_t, typename dst_t>
-static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
-    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void convert_unary(
+        const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+        const int64_t s01, const int64_t s02, const int64_t s03) {
+    const int64_t i00 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (i >= k) {
+    if (i00 >= ne00) {
         return;
     }
 
+    const int64_t i01 = blockIdx.y;
+    const int64_t i02 = blockIdx.z % ne02;
+    const int64_t i03 = blockIdx.z / ne02;
+
     const src_t * x = (const src_t *) vx;
 
-    y[i] = float(x[i]);
+    const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
+    const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
+    y[iy] = float(x[ix]);
 }
 
 template <typename src_t, typename dst_t>
-static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+static void convert_unary_cuda(const void * vx, dst_t * y,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t s01, const int64_t s02, const int64_t s03, cudaStream_t stream) {
+    const dim3 num_blocks((ne00 + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE, ne01, ne02*ne03);
+    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>
+        (vx, y, ne00, ne01, ne02, s01, s02, s03);
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_cont_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    convert_unary_cuda<src_t>(vx, y, k, 1, 1, 1, k, k, k, stream);
 }
 
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
         case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
         default:
             return nullptr;
     }
@@ -643,9 +661,9 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F32:
-            return convert_unary_cuda<float>;
+            return convert_unary_cont_cuda<float>;
         case GGML_TYPE_BF16:
-            return convert_unary_cuda<nv_bfloat16>;
+            return convert_unary_cont_cuda<nv_bfloat16>;
         default:
             return nullptr;
     }
@@ -692,7 +710,18 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
         case GGML_TYPE_IQ3_S:
             return dequantize_row_iq3_s_cuda;
         case GGML_TYPE_F16:
-            return convert_unary_cuda<half>;
+            return convert_unary_cont_cuda<half>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cont_cuda<nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
         case GGML_TYPE_BF16:
             return convert_unary_cuda<nv_bfloat16>;
         default:

ggml/src/ggml-cuda/convert.cuh

@@ -3,7 +3,7 @@
 #define CUDA_DEQUANTIZE_BLOCK_SIZE 256
 
 template<typename T>
-using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
+using to_t_cuda_t = void (*)(const void * x, T * y, int64_t k, cudaStream_t stream);
 
 typedef to_t_cuda_t<float> to_fp32_cuda_t;
 typedef to_t_cuda_t<half> to_fp16_cuda_t;
@@ -14,3 +14,13 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
 to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type);
 
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
+
+// TODO more general support for non-contiguous inputs
+
+template<typename T>
+using to_t_nc_cuda_t = void (*)(const void * x, T * y,
+    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
+    int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
+
+typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);

ggml/src/ggml-cuda/cpy.cu

@@ -592,6 +592,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
@@ -639,6 +641,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
     }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
 
 }

ggml/src/ggml-cuda/getrows.cu

@@ -33,8 +33,8 @@ static __global__ void k_get_rows(
     dfloat2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = v.x;
-    dst_row[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = float(v.x);
+    dst_row[iybs + iqs + y_offset] = float(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -60,7 +60,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = src0_row[i00];
+    dst_row[i00] = float(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>
@@ -86,120 +86,159 @@ static __global__ void k_get_rows_back_float(
     dst[dst_row*ncols + col] = sum;
 }
 
-template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
+template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
+static void get_rows_cuda_q(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     GGML_ASSERT(ne00 % 2 == 0);
 
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
-
-    GGML_UNUSED(dst);
 }
 
-template<typename src0_t>
+template<typename src0_t, typename dst_t>
 static void get_rows_cuda_float(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne13 == 1);
-
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
+}
 
-    GGML_UNUSED(dst);
+template <typename dst_t>
+static void ggml_cuda_get_rows_switch_src0_type(
+        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    switch (src0_type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
+            break;
+    }
+}
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream) {
+    switch (dst_type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
+            break;
+    }
 }
 
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const void    * src0_d = (const void    *) src0->data;
-    const int32_t * src1_d = (const int32_t *) src1->data;
-    float         * dst_d  = (float         *) dst->data;
-
     cudaStream_t stream = ctx.stream();
 
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ne13 == 1);
 
     GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
     GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
     GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));
 
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            break;
-    }
+    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
+        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
 }
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/getrows.cuh

@@ -3,6 +3,13 @@
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
 
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream);
+
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1410,6 +1410,11 @@ static void ggml_cuda_op_mul_mat(
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
+    // const int64_t nb10 = src1->nb[0];
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2];
+    const int64_t nb13 = src1->nb[3];
+
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
@@ -1545,7 +1550,10 @@ static void ggml_cuda_op_mul_mat(
             dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);
 
             if (src1_on_device && src1_is_contiguous) {
-                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
+                quantize_src1(
+                    dev[id].src1_ddf, nullptr, dev[id].src1_ddq, src0->type, ne10,
+                    nb11/sizeof(float), nb12/sizeof(float), nb13/sizeof(float),
+                    src1_padded_col_size, ne11, ne12, ne13, stream);
                 CUDA_CHECK(cudaGetLastError());
             }
         }
@@ -1640,7 +1648,9 @@ static void ggml_cuda_op_mul_mat(
                 }
 
                 if (quantize_src1 && !src1_is_contiguous) {
-                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
+                    quantize_src1(
+                        src1_ddf_i, nullptr, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
+                        src1_padded_col_size, src1_ncols, 1, 1, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
 
@@ -1710,15 +1720,15 @@ static __global__ void k_compute_batched_ptrs(
         size_t  nb12, size_t  nb13,
         size_t  nbd2, size_t  nbd3,
         int64_t r2,   int64_t r3) {
-    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
+    const int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    const int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
 
     if (i13 >= ne13 || i12 >= ne12) {
         return;
     }
 
-    int64_t i03 = i13 / r3;
-    int64_t i02 = i12 / r2;
+    const int64_t i03 = i13 / r3;
+    const int64_t i02 = i12 / r2;
 
     ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
     ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
@@ -1732,6 +1742,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
+    // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
+    // As long as dst is contiguous this does not matter though.
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t ne_dst = ggml_nelements(dst);
@@ -1740,21 +1754,31 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 
     CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
 
-    void * src0_ddq = src0->data;
-    half * src0_f16 = (half *) src0_ddq;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
+    const half * src0_f16 = (const half *) src0->data;
+    float * dst_ddf = (float *) dst->data;
+
+    const half * src1_f16 = (const half *) src1->data;
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    GGML_ASSERT(nb10 == ts_src1);
+    int64_t s11 = nb11 / ts_src1;
+    int64_t s12 = nb12 / ts_src1;
+    int64_t s13 = nb13 / ts_src1;
+    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
 
     // convert src1 to fp16
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
     if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
         const int64_t ne_src1 = ggml_nelements(src1);
         src1_f16_alloc.alloc(ne_src1);
         GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+
+        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+
+        src1_f16 = src1_f16_alloc.get();
+        s11 = ne10;
+        s12 = ne11*s11;
+        s13 = ne12*s12;
     }
-    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
 
     ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
     char * dst_t;
@@ -1814,13 +1838,13 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 int i02 = i12 / r2;
 
                 CUBLAS_CHECK(
-                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+                    ne01, ne11, ne10,
+                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
+                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
+                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
+                    cu_compute_type,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
             }
         }
     }
@@ -1831,15 +1855,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
-                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
-                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
+                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
+                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     } else {
         // use cublasGemmBatchedEx
-        const int ne23 = ne12*ne13;
+        const int64_t ne23 = ne12*ne13;
 
         ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
         ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
@@ -1851,8 +1875,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
-                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
+                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
+                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
                 nbd2, nbd3,
                 r2, r3);
         CUDA_CHECK(cudaGetLastError());
@@ -1861,8 +1885,8 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
                 alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
+                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
                 ne23,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1878,7 +1902,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
-    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
+    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
     bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
@@ -1919,12 +1943,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && use_mul_mat_vec && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
-        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
-               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_vec_q) {
+        ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_q) {
+        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
+            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
@@ -1938,196 +1966,145 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
-struct mmid_row_mapping {
-    int32_t i1;
-    int32_t i2;
-};
-
-static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
-                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
-                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
-                                                 int64_t ne11, int64_t ne10,
-                                                 size_t nb11, size_t nb12) {
-    int32_t iid1 = blockIdx.x;
-    int32_t id = blockIdx.y;
-
-    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
-
-    if (row_id_i != i02) {
-        return;
-    }
-
-    const int64_t i11 = id % ne11;
-    const int64_t i12 = iid1;
-
-    __shared__ int src1_row;
-    if (threadIdx.x == 0) {
-        src1_row = atomicAdd(cur_src1_row, 1);
-        row_mapping[src1_row] = {id, iid1};
-    }
-    __syncthreads();
-
-    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
-    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
-
-    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
-        src1_row_contiguous[i] = src1_row_original[i];
-    }
-}
-
-static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
-                                                  const mmid_row_mapping * __restrict__ row_mapping,
-                                                  int64_t ne0,
-                                                  size_t nb1, size_t nb2) {
-    int32_t i = blockIdx.x;
-
-    const int32_t i1 = row_mapping[i].i1;
-    const int32_t i2 = row_mapping[i].i2;
-
-    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
-    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
-
-    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
-        dst_row_original[j] = dst_row_contiguous[j];
-    }
-}
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * ids  = dst->src[2];
 
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (ne2 == 1) {
+            if (ggml_is_quantized(src0->type)) {
+                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+            } else {
+                ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
+            }
+            return;
+        }
+
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+            ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
+            return;
+        }
+    }
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t n_as = ne02;
-    const int64_t n_ids = ids->ne[0];
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc))
+        || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type;
+    const ggml_type type_dst_sorted  = GGML_TYPE_F32;
+    const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted);
+    const size_t ts_dst_sorted  = ggml_type_size(type_dst_sorted);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<int32_t> ids_to_sorted_host;
+    ids_to_sorted_host.reserve(2*ne_get_rows);
+    std::vector<int32_t> ids_from_sorted_host(ne_get_rows);
+
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool(), 2*ne_get_rows);
+
+    std::vector<int32_t> tokens_per_expert(ne02);
+
+    ggml_cuda_pool_alloc<char> src1_sorted(ctx.pool(), ne12*n_expert_used*ne10*ts_src1_sorted);
+    ggml_cuda_pool_alloc<char>  dst_sorted(ctx.pool(), ne2 *n_expert_used* ne0*ts_dst_sorted);
 
     std::vector<char> ids_host(ggml_nbytes(ids));
-    const char * ids_dev = (const char *) ids->data;
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
     CUDA_CHECK(cudaStreamSynchronize(stream));
 
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
-
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *)  dst->data;
-
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[3] = nb02;
-
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    if (ne12 == 1) {
-        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-            for (int64_t id = 0; id < n_ids; id++) {
-                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-                const int64_t i11 = id % ne11;
-                const int64_t i12 = iid1;
-
-                const int64_t i1 = id;
-                const int64_t i2 = i12;
-
-                src0_row.data = src0_original + i02*nb02;
-                src1_row.data = src1_original + i11*nb11 + i12*nb12;
-                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;
-
-                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-            }
-        }
-    } else {
-        ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
-        ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
-
-        src1_row.data = src1_contiguous.get();
-        dst_row.data  =  dst_contiguous.get();
-
-        for (int64_t i02 = 0; i02 < n_as; i02++) {
-            int64_t num_src1_rows = 0;
-
-            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-                for (int64_t id = 0; id < n_ids; id++) {
-                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
-                    if (row_id_i != i02) {
-                        continue;
-                    }
-
-                    num_src1_rows++;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_from_sorted_host[i12*n_expert_used + iex] = ids_to_sorted_host.size();
+                    ids_to_sorted_host.push_back(i12*ne11 + iex % ne11);
+                    tokens_per_expert[i02]++;
+                    break;
                 }
             }
-
-            if (num_src1_rows == 0) {
-                continue;
-            }
-
-            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
-            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
-            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
-
-            {
-                dim3 block_dims(std::min((unsigned int)ne10, 768u));
-                dim3 grid_dims(ids->ne[1], n_ids);
-                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        src1_original, src1_contiguous.get(),
-                        dev_cur_src1_row.get(), dev_row_mapping.get(),
-                        ids_dev, i02, ids->nb[1], ids->nb[0],
-                        ne11, ne10,
-                        nb11, nb12);
-                CUDA_CHECK(cudaGetLastError());
-            }
-
-            src0_row.data = src0_original + i02*nb02;
-
-            GGML_ASSERT(nb11 == sizeof(float)*ne10);
-            GGML_ASSERT(nb1 == sizeof(float)*ne0);
-
-            src1_row.ne[1] = num_src1_rows;
-            src1_row.nb[1] = nb11;
-            src1_row.nb[2] = num_src1_rows*nb11;
-            src1_row.nb[3] = num_src1_rows*nb11;
-
-            dst_row.ne[1] = num_src1_rows;
-            dst_row.nb[1] = nb1;
-            dst_row.nb[2] = num_src1_rows*nb1;
-            dst_row.nb[3] = num_src1_rows*nb1;
-
-            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-
-            {
-                dim3 block_dims(std::min((unsigned int)ne0, 768u));
-                dim3 grid_dims(num_src1_rows);
-                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        dst_original, dst_contiguous.get(),
-                        dev_row_mapping.get(),
-                        ne0,
-                        nb1, nb2);
-                CUDA_CHECK(cudaGetLastError());
-            }
         }
     }
+    GGML_ASSERT(ids_to_sorted_host.size() == size_t(ne_get_rows));
+
+    ids_to_sorted_host.insert(ids_to_sorted_host.end(), ids_from_sorted_host.begin(), ids_from_sorted_host.end());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_to_sorted_host.data(), 2*ne_get_rows*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_to_sorted   = ids_buf_dev.ptr + 0*ne_get_rows;
+    const int32_t * ids_from_sorted = ids_buf_dev.ptr + 1*ne_get_rows;
+
+    get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted,
+        ne10, nb11, nb12, nb13,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    char * src1_data_cur = (char *) src1_sorted.ptr;
+    char *  dst_data_cur = (char *)  dst_sorted.ptr;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) {
+        if (tokens_per_expert[i02] == 0) {
+            continue;
+        }
+
+        ggml_tensor src0_slice = *src0;
+        src0_slice.ne[2] = 1;
+        src0_slice.nb[3] = src0_slice.nb[2];
+        src0_slice.data  = (char *) src0->data + i02*nb02;
+
+        ggml_tensor src1_slice;
+        memset(&src1_slice, 0, sizeof(src1_slice));
+        src1_slice.buffer = src1->buffer;
+        src1_slice.type   = type_src1_sorted;
+        src1_slice.ne[0]  = ne10;
+        src1_slice.ne[1]  = tokens_per_expert[i02];
+        src1_slice.ne[2]  = 1;
+        src1_slice.ne[3]  = 1;
+        src1_slice.nb[0]  = ts_src1_sorted;
+        src1_slice.nb[1]  = src1_slice.ne[0] * src1_slice.nb[0];
+        src1_slice.nb[2]  = src1_slice.ne[1] * src1_slice.nb[1];
+        src1_slice.nb[3]  = src1_slice.ne[2] * src1_slice.nb[2];
+        src1_slice.data   = src1_data_cur;
+
+        ggml_tensor dst_slice;
+        memset(&dst_slice, 0, sizeof(dst_slice));
+        dst_slice.buffer = dst->buffer;
+        dst_slice.type   = type_dst_sorted;
+        dst_slice.ne[0]  = ne0;
+        dst_slice.ne[1]  = tokens_per_expert[i02];
+        dst_slice.ne[2]  = 1;
+        dst_slice.ne[3]  = 1;
+        dst_slice.nb[0]  = ts_dst_sorted;
+        dst_slice.nb[1]  = dst_slice.ne[0] * dst_slice.nb[0];
+        dst_slice.nb[2]  = dst_slice.ne[1] * dst_slice.nb[1];
+        dst_slice.nb[3]  = dst_slice.ne[2] * dst_slice.nb[2];
+        dst_slice.data   = dst_data_cur;
+
+        ggml_cuda_mul_mat(ctx, &src0_slice, &src1_slice, &dst_slice);
+        CUDA_CHECK(cudaGetLastError());
+
+        src1_data_cur += src1_slice.nb[2];
+        dst_data_cur  +=  dst_slice.nb[2];
+    }
+
+    get_rows_cuda(dst_sorted.ptr, type_dst_sorted, ids_from_sorted, dst->data, dst->type,
+        ne0, ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        nb1, nb2, nb3, stream);
 }
 
 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
@@ -2489,7 +2466,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_MUL_MAT_ID) {
+        if (node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
             use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
@@ -3203,9 +3180,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         }
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK: {
-            const size_t ts = ggml_type_size(op->src[0]->type);
-            const int64_t ne0_012 = op->src[0]->ne[0] * op->src[0]->ne[1] * op->src[0]->ne[2];
-            return op->src[0]->nb[0] == ts && op->src[0]->nb[3] == ne0_012*ts;
+            return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
         }
         case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:

ggml/src/ggml-cuda/mmq.cu

@@ -1,37 +1,10 @@
 #include "mmq.cuh"
+#include "quantize.cuh"
 
-void ggml_cuda_op_mul_mat_q(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
+#include <vector>
 
-    const int64_t ne00 = src0->ne[0];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    GGML_ASSERT(ne10 % QK8_1 == 0);
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
-
-    int id = ggml_cuda_get_device();
-    const int cc = ggml_cuda_info().devices[id].cc;
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
-
-    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
-    // Also its fixup needs to allocate a temporary buffer in the memory pool.
-    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
-
-    switch (src0->type) {
+static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    switch (args.type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -90,10 +63,195 @@ void ggml_cuda_op_mul_mat_q(
             GGML_ABORT("fatal error");
             break;
     }
+}
+
+void ggml_cuda_mul_mat_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    cudaStream_t stream = ctx.stream();
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
+
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+
+    const char  * src0_d = (const char  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
+
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
+
+    if (!ids) {
+        const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+            get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+        ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+        {
+            const int64_t s11 = src1->nb[1] / ts_src1;
+            const int64_t s12 = src1->nb[2] / ts_src1;
+            const int64_t s13 = src1->nb[3] / ts_src1;
+            quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
+                ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        }
+
+        const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+        const int64_t s13 = ne12*s12;
+
+        const mmq_args args = {
+            src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
+            ne00, ne01, ne1, s01, s1,
+            ne02, ne12, s02, s12, s2,
+            ne03, ne13, s03, s13, s3,
+            use_stream_k};
+        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+        return;
+    }
+
+    GGML_ASSERT(ne13 == 1);
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    std::vector<int32_t> ids_src1_host;
+    ids_src1_host.reserve(ne_get_rows);
+    std::vector<int32_t> ids_dst_host;
+    ids_dst_host.reserve(ne_get_rows);
+    std::vector<int32_t> tokens_per_expert_host(ne02);
+    std::vector<int32_t> expert_bounds_host(ne02 + 1);
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
+                    ids_dst_host.push_back(i12*ne1 + iex);
+                    tokens_per_expert_host[i02]++;
+                    break;
+                }
+            }
+        }
+    }
+
+    int32_t cumsum = 0;
+    for (int64_t i = 0; i < ne02; ++i) {
+        expert_bounds_host[i] = cumsum;
+        cumsum += tokens_per_expert_host[i];
+    }
+    expert_bounds_host[ne02] = cumsum;
+
+    std::vector<int32_t> ids_buf_host;
+    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
+    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
+    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
+    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
+    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
+
+    const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+        get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+    const int64_t ne11_flat = ne12*n_expert_used;
+    const int64_t ne12_flat = 1;
+    const int64_t ne13_flat = 1;
+
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[2] / ts_src1;
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+            ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+    }
+
+    const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+    const int64_t s13 = ne12*s12;
+
+    // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
+    const mmq_args args = {
+        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        ne00, ne01, ne_get_rows, s01, s1,
+        ne02, ne02, s02, s12, s2,
+        ne03, ne13, s03, s13, s3,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+}
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride01 = ne00 / ggml_blck_size(src0->type);
+
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
+    const mmq_args args = {
+        src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
+        ne00, row_diff, src1_ncols, stride01, nrows_dst,
+        1, 1, 0, 0, 0,
+        1, 1, 0, 0, 0,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {

ggml/src/ggml-cuda/mmv.cu

@@ -4,18 +4,23 @@
 
 template <typename T, typename type_acc, int block_size>
 static __global__ void mul_mat_vec(
-        const T * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row,
         const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
         const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst) {
-    const int64_t row       = blockIdx.x;
-    const int64_t channel   = blockIdx.y;
-    const int64_t sample    = blockIdx.z;
-    const int     tid       = threadIdx.x;
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int64_t row         = blockIdx.x;
+    const int64_t channel_dst = blockIdx.y;
+    const int64_t channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int64_t channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
+    const int64_t sample_dst  = blockIdx.z;
+    const int64_t sample_x    = sample_dst / sample_ratio;
+    const int64_t sample_y    = sample_dst;
+    const int     tid         = threadIdx.x;
+    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
 
-    x   +=  (sample/sample_ratio)*stride_sample_x   + (channel/channel_ratio)*stride_channel_x + row*stride_row;
-    y   +=   sample              *stride_sample_y   +  channel               *stride_channel_y;
-    dst +=   sample              *stride_sample_dst +  channel               *stride_channel_dst;
+    x   += sample_x  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    y   += sample_y  *stride_sample_y   + channel_y  *stride_channel_y;
+    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst;
 
     const float2 * y2 = (const float2 *) y;
 
@@ -31,12 +36,19 @@ static __global__ void mul_mat_vec(
 
     float sumf = 0.0f;
 
-    if constexpr (std::is_same<T, half>::value) {
+    if constexpr (std::is_same<T, float>::value) {
+        const float2 * x2 = (const float2 *) x;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = x2[col2];
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x*tmpy.x;
+            sumf += tmpx.y*tmpy.y;
+        }
+    } else if constexpr (std::is_same<T, half>::value) {
         const half2 * x2 = (const half2 *) x;
 
         if (std::is_same<type_acc, float>::value) {
-            sumf = 0.0f;
-
             for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
                 const float2 tmpy = y2[col2];
@@ -59,8 +71,6 @@ static __global__ void mul_mat_vec(
         }
     } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
         const int * x2 = (const int *) x;
-        sumf = 0.0f;
-
         for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
             const int    tmpx = x2[col2];
             const float2 tmpy = y2[col2];
@@ -92,17 +102,17 @@ static __global__ void mul_mat_vec(
 
 template <typename T, typename type_acc>
 static void launch_mul_mat_vec_cuda(
-        const T * x, const float * y, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
-        const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream) {
     GGML_ASSERT(ncols      % 2 == 0);
     GGML_ASSERT(stride_row % 2 == 0);
-    GGML_ASSERT(nchannels_y % nchannels_x == 0);
-    GGML_ASSERT(nsamples_y  % nsamples_x  == 0);
-    const int64_t channel_ratio = nchannels_y / nchannels_x;
-    const int64_t sample_ratio  = nsamples_y  / nsamples_x;
+    GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
+    GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const int64_t channel_ratio = nchannels_dst / nchannels_x;
+    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
     int device;
     int warp_size;
 
@@ -124,48 +134,48 @@ static void launch_mul_mat_vec_cuda(
     }
 
     const int smem = warp_size*sizeof(float);
-    const dim3 block_nums(nrows, nchannels_y, nsamples_y);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
             mul_mat_vec<T, type_acc,  32><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case   64: {
             mul_mat_vec<T, type_acc,  64><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case   96: {
             mul_mat_vec<T, type_acc,  96><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  128: {
             mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  160: {
             mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  192: {
             mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  224: {
             mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  256: {
             mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -175,28 +185,28 @@ static void launch_mul_mat_vec_cuda(
 
 template<typename T>
 static void mul_mat_vec_cuda(
-        const T * x, const float * y, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
-        const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         enum ggml_prec prec, cudaStream_t stream) {
-    switch (prec) {
-        case GGML_PREC_DEFAULT: {
+    if constexpr(std::is_same<T, half>::value) {
+        if (prec == GGML_PREC_DEFAULT) {
             launch_mul_mat_vec_cuda<T, half>
-                (x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
-        } break;
-        case GGML_PREC_F32: {
-            launch_mul_mat_vec_cuda<T, float>
-                (x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
-        } break;
+                (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            return;
+        }
     }
+    launch_mul_mat_vec_cuda<T, float>
+        (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+         stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
 }
 
-void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(         dst->type == GGML_TYPE_F32);
 
     GGML_TENSOR_BINARY_OP_LOCALS;
 
@@ -204,21 +214,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
     const size_t ts_src1 = ggml_type_size(src1->type);
     const size_t ts_dst  = ggml_type_size(dst->type);
 
-    GGML_ASSERT(ne11 == 1);
-    GGML_ASSERT(ne12 == ne2);
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
     GGML_ASSERT(ne13 == ne3);
 
-    GGML_ASSERT(nb00 == ts_src0);
-    GGML_ASSERT(nb10 == ts_src1);
-    GGML_ASSERT(nb0  == ts_dst);
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+    GGML_ASSERT(        nb0        == ts_dst);
 
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
 
-    const float * src1_d = (const float *) src1->data;
-    float       *  dst_d = (float       *)  dst->data;
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;
 
     const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = src1->nb[1] / ts_src1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
     const int64_t s02 = src0->nb[2] / ts_src0;
     const int64_t s12 = src1->nb[2] / ts_src1;
     const int64_t s2  =  dst->nb[2] / ts_dst;
@@ -226,14 +239,33 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
     const int64_t s13 = src1->nb[3] / ts_src1;
     const int64_t s3  =  dst->nb[3] / ts_dst;
 
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
+
+    GGML_ASSERT(ncols_dst == 1);
+
     switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0->data;
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+        } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -262,27 +294,34 @@ void ggml_cuda_op_mul_mat_vec(
     const int64_t stride_row         = ne00;
     const int64_t nchannels_x        = 1;
     const int64_t nchannels_y        = 1;
+    const int64_t nchannels_dst      = 1;
     const int64_t stride_channel_x   = 0;
     const int64_t stride_channel_y   = 0;
     const int64_t stride_channel_dst = 0;
     const int64_t nsamples_x         = 1;
-    const int64_t nsamples_y         = 1;
+    const int64_t nsamples_dst       = 1;
     const int64_t stride_sample_x    = 0;
     const int64_t stride_sample_y    = 0;
     const int64_t stride_sample_dst  = 0;
 
     switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0_dd_i;
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+        } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
-                nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
-                nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));

ggml/src/ggml-cuda/mmv.cuh

@@ -3,7 +3,7 @@
 // maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
 #define MMV_MAX_ROWS 512
 
-void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
 void ggml_cuda_op_mul_mat_vec(
     ggml_backend_cuda_context & ctx,

ggml/src/ggml-cuda/mmvq.cu

@@ -1,50 +1,57 @@
 #include "mmvq.cuh"
+#include "quantize.cuh"
 #include "vecdotq.cuh"
 
+#include <cstdint>
+
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
 
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? vec_dot_q4_0_q8_1 :
-        type == GGML_TYPE_Q4_1 ? vec_dot_q4_1_q8_1 :
-        type == GGML_TYPE_Q5_0 ? vec_dot_q5_0_q8_1 :
-        type == GGML_TYPE_Q5_1 ? vec_dot_q5_1_q8_1 :
-        type == GGML_TYPE_Q8_0 ? vec_dot_q8_0_q8_1 :
-        type == GGML_TYPE_Q2_K ? vec_dot_q2_K_q8_1 :
-        type == GGML_TYPE_Q3_K ? vec_dot_q3_K_q8_1 :
-        type == GGML_TYPE_Q4_K ? vec_dot_q4_K_q8_1 :
-        type == GGML_TYPE_Q5_K ? vec_dot_q5_K_q8_1 :
-        type == GGML_TYPE_Q6_K ? vec_dot_q6_K_q8_1 :
-        type == GGML_TYPE_IQ2_XXS ? vec_dot_iq2_xxs_q8_1 :
-        type == GGML_TYPE_IQ2_XS ? vec_dot_iq2_xs_q8_1 :
-        type == GGML_TYPE_IQ2_S ? vec_dot_iq2_s_q8_1 :
-        type == GGML_TYPE_IQ3_XXS ? vec_dot_iq3_xxs_q8_1 :
-        type == GGML_TYPE_IQ1_S ? vec_dot_iq1_s_q8_1 :
-        type == GGML_TYPE_IQ1_M ? vec_dot_iq1_m_q8_1 :
-        type == GGML_TYPE_IQ4_NL ? vec_dot_iq4_nl_q8_1 :
-        type == GGML_TYPE_IQ4_XS ? vec_dot_iq4_xs_q8_1 :
-        type == GGML_TYPE_IQ3_S ? vec_dot_iq3_s_q8_1 :
-        nullptr;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
+        case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
+        case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
+        case GGML_TYPE_Q5_1:    return vec_dot_q5_1_q8_1;
+        case GGML_TYPE_Q8_0:    return vec_dot_q8_0_q8_1;
+        case GGML_TYPE_Q2_K:    return vec_dot_q2_K_q8_1;
+        case GGML_TYPE_Q3_K:    return vec_dot_q3_K_q8_1;
+        case GGML_TYPE_Q4_K:    return vec_dot_q4_K_q8_1;
+        case GGML_TYPE_Q5_K:    return vec_dot_q5_K_q8_1;
+        case GGML_TYPE_Q6_K:    return vec_dot_q6_K_q8_1;
+        case GGML_TYPE_IQ2_XXS: return vec_dot_iq2_xxs_q8_1;
+        case GGML_TYPE_IQ2_XS:  return vec_dot_iq2_xs_q8_1;
+        case GGML_TYPE_IQ2_S:   return vec_dot_iq2_s_q8_1;
+        case GGML_TYPE_IQ3_XXS: return vec_dot_iq3_xxs_q8_1;
+        case GGML_TYPE_IQ1_S:   return vec_dot_iq1_s_q8_1;
+        case GGML_TYPE_IQ1_M:   return vec_dot_iq1_m_q8_1;
+        case GGML_TYPE_IQ4_NL:  return vec_dot_iq4_nl_q8_1;
+        case GGML_TYPE_IQ4_XS:  return vec_dot_iq4_xs_q8_1;
+        case GGML_TYPE_IQ3_S:   return vec_dot_iq3_s_q8_1;
+        default:                return nullptr;
+    }
 }
 
 static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q4_1    ? VDR_Q4_1_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_0    ? VDR_Q5_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_1    ? VDR_Q5_1_Q8_1_MMVQ :
-        type == GGML_TYPE_Q8_0    ? VDR_Q8_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q2_K    ? VDR_Q2_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q3_K    ? VDR_Q3_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q4_K    ? VDR_Q4_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_K    ? VDR_Q5_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q6_K    ? VDR_Q6_K_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_XS  ? VDR_IQ2_XS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_S   ? VDR_IQ2_S_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ3_S   ? VDR_IQ3_S_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ4_NL  ? VDR_IQ4_NL_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ4_XS  ? VDR_IQ4_XS_Q8_1_MMVQ :
-        1;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_1:    return VDR_Q5_1_Q8_1_MMVQ;
+        case GGML_TYPE_Q8_0:    return VDR_Q8_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q2_K:    return VDR_Q2_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q3_K:    return VDR_Q3_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q4_K:    return VDR_Q4_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_K:    return VDR_Q5_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q6_K:    return VDR_Q6_K_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_XXS: return VDR_IQ2_XXS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_XS:  return VDR_IQ2_XS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_S:   return VDR_IQ2_S_Q8_1_MMVQ;
+        case GGML_TYPE_IQ3_XXS: return VDR_IQ3_XXS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ3_S:   return VDR_IQ3_S_Q8_1_MMVQ;
+        case GGML_TYPE_IQ4_NL:  return VDR_IQ4_NL_Q8_1_MMVQ;
+        case GGML_TYPE_IQ4_XS:  return VDR_IQ4_XS_Q8_1_MMVQ;
+        default:                return 1;
+    }
 }
 
 enum mmvq_parameter_table_id {
@@ -73,9 +80,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
-static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(int ncols_dst,  mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
             case 1:
             case 2:
             case 3:
@@ -90,7 +97,7 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_paramete
                 return 1;
         }
     } else if (table_id == MMVQ_PARAMETERS_GCN) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
             case 1:
             case 2:
             case 3:
@@ -107,9 +114,9 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_paramete
     return 1;
 }
 
-static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) {
+static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
             case 1:
                 return 1;
             case 2:
@@ -127,19 +134,21 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int ta
     return 1;
 }
 
-template <ggml_type type, int ncols_y>
+template <ggml_type type, int ncols_dst>
 // tell the compiler to use as many registers as it wants, see nwarps definition below
-__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int ncols_x, const int nchannels_y, const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
     constexpr int vdr = get_vdr_mmvq(type);
     constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(ncols_y, table_id);
-    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id);
+    constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
+    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
@@ -147,13 +156,21 @@ static __global__ void mul_mat_vec_q(
     const     int tid = warp_size*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;
     const     int blocks_per_row_x = ncols_x / qk;
-    const     int blocks_per_col_y = nrows_y / QK8_1;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // partial sum for each thread
-    float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}};
+    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
+    const int channel_dst = blockIdx.y;
+    const int channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_y   = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
 
-    const block_q8_1 * y = (const block_q8_1 *) vy;
+    // partial sum for each thread
+    float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
+    const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
         const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
@@ -162,18 +179,19 @@ static __global__ void mul_mat_vec_q(
         const int kqs = vdr * (tid % (qi/vdr));
 
 #pragma unroll
-        for (int j = 0; j < ncols_y; ++j) {
+        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
-                tmp[j][i] += vec_dot_q_cuda(vx, &y[j*blocks_per_col_y + kby], (row0 + i)*blocks_per_row_x + kbx, kqs);
+                tmp[j][i] += vec_dot_q_cuda(
+                    vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
             }
         }
     }
 
-    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size];
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
     if (threadIdx.y > 0) {
 #pragma unroll
-        for (int j = 0; j < ncols_y; ++j) {
+        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
@@ -185,9 +203,11 @@ static __global__ void mul_mat_vec_q(
         return;
     }
 
+    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
+
     // sum up partial sums and write back result
 #pragma unroll
-    for (int j = 0; j < ncols_y; ++j) {
+    for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
         for (int i = 0; i < rows_per_cuda_block; ++i) {
 #pragma unroll
@@ -197,88 +217,121 @@ static __global__ void mul_mat_vec_q(
             tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
         }
 
-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) {
-            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + int(threadIdx.x) < stride_col_dst)) {
+            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }
-
-    GGML_UNUSED(nrows_x);
 }
 
-static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {
-    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id);
-    const dim3 block_nums(nblocks, 1, 1);
-    const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1);
+static std::pair<dim3, dim3> calc_launch_params(
+        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
+        const int warp_size, const mmvq_parameter_table_id table_id) {
+    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
+    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
+    const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
     return {block_nums, block_dims};
 }
 
 template <ggml_type type>
-static void mul_mat_vec_q_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+static void mul_mat_vec_q_switch_ncols_dst(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst,
+        const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
+        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
-    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+    GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
+
+    const int channel_ratio = nchannels_dst / nchannels_x;
+    const int sample_ratio  = nsamples_dst  / nsamples_x;
 
     const int device = ggml_cuda_get_device();
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
-    switch (ncols_y) {
+    GGML_ASSERT(!ids || ncols_dst == 1);
+    switch (ncols_dst) {
         case 1:
         {
-            constexpr int c_ncols_y = 1;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 1;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 2:
         {
-            constexpr int c_ncols_y = 2;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 2;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 3:
         {
-            constexpr int c_ncols_y = 3;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 3;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 4:
         {
-            constexpr int c_ncols_y = 4;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 4;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 5:
         {
-            constexpr int c_ncols_y = 5;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 5;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 6:
         {
-            constexpr int c_ncols_y = 6;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 6;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 7:
         {
-            constexpr int c_ncols_y = 7;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 7;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 8:
         {
-            constexpr int c_ncols_y = 8;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 8;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         default:
@@ -287,137 +340,213 @@ static void mul_mat_vec_q_cuda(
     }
 }
 
-static void mul_mat_vec_q4_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+static void mul_mat_vec_q_switch_type(
+        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst,
+        const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
+        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        cudaStream_t stream) {
+    switch (type_x) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
 }
 
-static void mul_mat_vec_q4_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+void ggml_cuda_mul_mat_vec_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
-static void mul_mat_vec_q5_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    cudaStream_t stream = ctx.stream();
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
 
-static void mul_mat_vec_q5_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
 
-static void mul_mat_vec_q8_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q8_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1);
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[3] / ts_src1;
+        quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+    }
 
-static void mul_mat_vec_q2_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = ne10_padded / QK8_1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q2_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const int64_t s12 = ne11*s11;
+    const int64_t s13 = ne12*s12;
 
-static void mul_mat_vec_q3_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_col_dst     = ids ? s2   : s1;
+    const int64_t stride_col_y       = ids ? s12  : s11;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q3_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q4_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q5_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q6_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q6_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_xxs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_xs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq3_xxs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq1_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq1_m_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_M>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq4_nl_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_NL>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq4_xs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq3_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+    mul_mat_vec_q_switch_type(
+        src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
+        ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
+        ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+        ne03,              ne3,           s03, s13,              s3,                 stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -440,68 +569,12 @@ void ggml_cuda_op_mul_mat_vec_q(
     // nrows_dst == nrows of the matrix that the kernel writes into
     const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
 
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_XXS:
-            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_XS:
-            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_S:
-            mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ3_XXS:
-            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ1_S:
-            mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ1_M:
-            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_NL:
-            mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_XS:
-            mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ3_S:
-            mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
+    const int stride_row_x = ne00 / ggml_blck_size(src0->type);
+    const int stride_col_y = src1_padded_row_size / QK8_1;
+
+    mul_mat_vec_q_switch_type(
+        src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);

ggml/src/ggml-cuda/mmvq.cuh

@@ -2,6 +2,9 @@
 
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
 
+void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
 void ggml_cuda_op_mul_mat_vec_q(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,

ggml/src/ggml-cuda/quantize.cu

@@ -1,30 +1,40 @@
 #include "quantize.cuh"
 #include <cstdint>
 
-static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx0_padded) {
-    const int64_t ix0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void quantize_q8_1(
+        const float * __restrict__ x, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {
+    const int64_t i0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
         return;
     }
 
-    const int64_t ix1 = blockIdx.y;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
 
-    const int64_t i_padded = ix1*kx0_padded + ix0;
+    const int64_t & i00 = i0;
+    const int64_t & i01 = i1;
+    const int64_t & i02 = i2;
+    const int64_t & i03 = i3;
+
+    const int64_t i_cont = ((i3*ne2 + i2) * ne1 + i1) * ne0 + i0;
 
     block_q8_1 * y = (block_q8_1 *) vy;
 
-    const int64_t ib = i_padded / QK8_1; // block index
-    const int64_t iqs = i_padded % QK8_1; // quant index
+    const int64_t ib  = i_cont / QK8_1; // block index
+    const int64_t iqs = i_cont % QK8_1; // quant index
 
-    const float xi = ix0 < kx ? x[ix1*kx + ix0] : 0.0f;
+    const float xi = i0 < ne00 ? x[i03*s03 + i02*s02 + i01*s01 + i00] : 0.0f;
     float amax = fabsf(xi);
     float sum = xi;
 
     amax = warp_reduce_max(amax);
-    sum = warp_reduce_sum(sum);
+    sum  = warp_reduce_sum(sum);
 
-    const float d = amax / 127;
+    const float  d = amax / 127;
     const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
 
     y[ib].qs[iqs] = q;
@@ -39,29 +49,38 @@ static __global__ void quantize_q8_1(const float * __restrict__ x, void * __rest
 
 template <mmq_q8_1_ds_layout ds_layout>
 static __global__ void quantize_mmq_q8_1(
-    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+        const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {
 
     constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
     constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
 
-    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
 
-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
         return;
     }
 
-    const float4 * x4 = (const float4 *) x;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
 
-    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
+    const int64_t i00 = i0;
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t i02 = i2;
+    const int64_t i03 = i3;
+
+    const float4 * x4 = (const float4 *) x;
 
     block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
 
     const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
-    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block
 
     // Load 4 floats per thread and calculate max. abs. value between them:
-    const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
     float amax = fabsf(xi.x);
     amax = fmaxf(amax, fabsf(xi.y));
     amax = fmaxf(amax, fabsf(xi.z));
@@ -77,7 +96,7 @@ static __global__ void quantize_mmq_q8_1(
     if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
         sum = xi.x + xi.y + xi.z + xi.w;
 
-        // Exchange calculate sum across vals_per_sum/4 threads.
+        // Calculate sums across vals_per_sum/4 threads.
 #pragma unroll
         for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
             sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
@@ -127,40 +146,40 @@ static __global__ void quantize_mmq_q8_1(
 }
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(!ids);
+    GGML_ASSERT(ne0 % QK8_1 == 0);
 
-    GGML_ASSERT(kx0_padded % QK8_1 == 0);
-
-    const int64_t block_num_x = (kx0_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, kx1*channels, 1);
+    const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx0_padded);
-
-    GGML_UNUSED(type_x);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    GGML_UNUSED(type_src0);
 }
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % (4*QK8_1) == 0);
 
-    GGML_ASSERT(kx0_padded % (4*QK8_1) == 0);
-
-    const int64_t block_num_x = (kx0_padded + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
-    const dim3 num_blocks(block_num_x, kx1, channels);
+    const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
-    switch (mmq_get_q8_1_ds_layout(type_x)) {
+    switch (mmq_get_q8_1_ds_layout(type_src0)) {
         case MMQ_Q8_1_DS_LAYOUT_D4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         case MMQ_Q8_1_DS_LAYOUT_DS4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         case MMQ_Q8_1_DS_LAYOUT_D2S6:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         default:
             GGML_ABORT("fatal error");

ggml/src/ggml-cuda/quantize.cuh

@@ -12,13 +12,16 @@ static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk
 static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");
 
 typedef void (*quantize_cuda_t)(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

ggml/src/ggml-cuda/vecdotq.cuh

@@ -1,3 +1,5 @@
+#pragma once
+
 #include "common.cuh"
 #include <cstdint>
 

ggml/src/ggml-metal/ggml-metal.m

@@ -44,8 +44,8 @@ static struct ggml_backend_device g_ggml_backend_metal_device;
 // note: assumes single GPU device - the default one
 // TODO: support multiple GPU devices
 static struct ggml_backend_metal_device_context {
-    id<MTLDevice> mtl_device;
-    int           mtl_device_ref_count;
+    id<MTLDevice>  mtl_device;
+    int            mtl_device_ref_count;
     id<MTLLibrary> mtl_library;
 
     bool has_simdgroup_reduction;
@@ -490,7 +490,259 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COUNT
 };
 
+//
+// ggml_metal_heap
+//
+
+struct ggml_metal_heap {
+    // number of times the heap was unused
+    int n_unused;
+
+    // total number of buffer allocations in this heap across all computes
+    int64_t n_alloc;
+
+    // current offset in the heap - we reset this after each node in order to reuse the memory
+    size_t offs;
+
+    // the currently allocated MTLBuffer objects in this heap
+    id<MTLHeap> obj;
+
+    NSMutableArray * bufs;
+};
+
+static struct ggml_metal_heap * ggml_metal_heap_init(id<MTLDevice> device, size_t size) {
+    struct ggml_metal_heap * heap = calloc(1, sizeof(struct ggml_metal_heap));
+
+    MTLHeapDescriptor * desc = [[MTLHeapDescriptor alloc] init];
+    desc.storageMode  = MTLStorageModePrivate;
+    desc.cpuCacheMode = MTLCPUCacheModeDefaultCache;
+    desc.type         = MTLHeapTypePlacement;
+    desc.size         = size;
+
+    heap->n_unused = 0;
+    heap->n_alloc = 0;
+
+    heap->obj = [device newHeapWithDescriptor:desc];
+    if (!heap->obj) {
+        GGML_LOG_ERROR("%s: error: failed to create MTLHeap with size %zu\n", __func__, size);
+
+        free(heap);
+
+        return false;
+    }
+
+    [desc release];
+
+    heap->bufs = [[NSMutableArray alloc] init];
+
+    return heap;
+}
+
+static void ggml_metal_heap_reset(struct ggml_metal_heap * heap) {
+    heap->offs = 0;
+
+    // count how many graph computes the heap ended up being unused
+    if ([heap->bufs count] > 0) {
+        heap->n_unused = 0;
+    } else {
+        heap->n_unused++;
+    }
+
+    for (id<MTLBuffer> buf in heap->bufs) {
+        [buf release];
+    }
+    [heap->bufs removeAllObjects];
+
+    // tell the OS that it can reuse this memory if needed
+    // ref: https://developer.apple.com/documentation/metal/mtlpurgeablestate?language=objc
+    [heap->obj setPurgeableState:MTLPurgeableStateVolatile];
+}
+
+static void ggml_metal_heap_free(struct ggml_metal_heap * heap) {
+    if (heap == nil) {
+        return;
+    }
+
+    ggml_metal_heap_reset(heap);
+
+    [heap->obj  release];
+    [heap->bufs release];
+
+    free(heap);
+}
+
+@interface ggml_metal_heap_ptr : NSObject
+
+@property (nonatomic, assign) struct ggml_metal_heap * data;
+
+@end
+
+@implementation ggml_metal_heap_ptr
+@end
+
+//
+// ggml_metal_mem_pool
+//
+
+struct ggml_metal_mem_pool {
+    id<MTLDevice> device;
+
+    int n_heaps; // total number of heaps ever created (including those that were removed)
+
+    NSMutableArray * heaps;
+    NSMutableArray * heaps_to_remove;
+};
+
+static struct ggml_metal_mem_pool * ggml_metal_mem_pool_init(void) {
+    struct ggml_metal_mem_pool * mem_pool = calloc(1, sizeof(struct ggml_metal_mem_pool));
+
+    mem_pool->n_heaps = 0;
+
+    mem_pool->heaps           = [[NSMutableArray alloc] init];
+    mem_pool->heaps_to_remove = [[NSMutableArray alloc] init];
+
+    return mem_pool;
+}
+
+static void ggml_metal_mem_pool_free(struct ggml_metal_mem_pool * mem_pool) {
+    GGML_LOG_DEBUG("%s: freeing memory pool, num heaps = %zu (total = %d)\n", __func__, [mem_pool->heaps count], mem_pool->n_heaps);
+
+    size_t size_all = 0;
+    size_t size_cur = 0;
+
+    for (ggml_metal_heap_ptr * ptr in mem_pool->heaps) {
+        GGML_LOG_DEBUG("%s:   heap: %p\n",                __func__, (void *) ptr.data);
+        GGML_LOG_DEBUG("%s:     n_alloc:  %" PRId64 "\n", __func__, ptr.data->n_alloc);
+        GGML_LOG_DEBUG("%s:     n_unused: %d\n",          __func__, ptr.data->n_unused);
+        GGML_LOG_DEBUG("%s:     size:     %.2f MiB\n",    __func__, [ptr.data->obj size] / 1024.0 / 1024.0);
+        GGML_LOG_DEBUG("%s:     bufs:     %zu\n",         __func__, [ptr.data->bufs count]);
+
+        if ([ptr.data->bufs count] > 0) {
+            size_cur += [ptr.data->obj size];
+        }
+        size_all += [ptr.data->obj size];
+
+        ggml_metal_heap_free(ptr.data);
+        [ptr release];
+    }
+    [mem_pool->heaps           release];
+    [mem_pool->heaps_to_remove release];
+
+    if (size_all > 0) {
+        GGML_LOG_DEBUG("%s:   size_all: %.2f MiB\n", __func__, size_all / 1024.0 / 1024.0);
+        GGML_LOG_DEBUG("%s:   size_cur: %.2f MiB\n", __func__, size_cur / 1024.0 / 1024.0);
+    }
+
+    free(mem_pool);
+}
+
+static void ggml_metal_mem_pool_reset(struct ggml_metal_mem_pool * mem_pool) {
+    for (NSUInteger i = 0; i < [mem_pool->heaps count]; i++) {
+        ggml_metal_heap_ptr * ptr = [mem_pool->heaps objectAtIndex:i];
+
+        struct ggml_metal_heap * heap = ptr.data;
+        ggml_metal_heap_reset(heap);
+
+        // if the heap hasn't been used for a while, remove it
+        if (heap->n_unused >= 128) {
+            [mem_pool->heaps_to_remove addObject:@(i)];
+        }
+    }
+
+    if (mem_pool->heaps_to_remove.count > 0) {
+        for (NSUInteger i = 0; i < [mem_pool->heaps_to_remove count]; i++) {
+            NSUInteger index = [[mem_pool->heaps_to_remove objectAtIndex:i] intValue];
+            ggml_metal_heap_ptr * ptr = [mem_pool->heaps objectAtIndex:index];
+
+            struct ggml_metal_heap * heap = ptr.data;
+            ggml_metal_heap_free(heap);
+
+            [mem_pool->heaps removeObjectAtIndex:index];
+            [ptr release];
+        }
+
+        [mem_pool->heaps_to_remove removeAllObjects];
+    }
+}
+
+static void ggml_metal_mem_pool_clear(struct ggml_metal_mem_pool * mem_pool) {
+    for (ggml_metal_heap_ptr * ptr in mem_pool->heaps) {
+        ptr.data->offs = 0;
+    }
+}
+
+static id<MTLBuffer> ggml_metal_mem_pool_alloc(struct ggml_metal_mem_pool * mem_pool, size_t size) {
+    const size_t alignment = 32;
+
+    const size_t size_aligned = GGML_PAD(size, alignment);
+
+    // try one of the existing heaps
+    for (ggml_metal_heap_ptr * ptr in mem_pool->heaps) {
+        struct ggml_metal_heap * heap = ptr.data;
+        if (heap->offs + size_aligned <= [heap->obj size]) {
+            // if this is the first buffer in the heap for the current command buffer, tell the OS that
+            //   it cannot free the memory used by the heap
+            // ref: https://developer.apple.com/documentation/metal/mtlpurgeablestate?language=objc
+            if ([heap->bufs count] == 0) {
+                [heap->obj setPurgeableState:MTLPurgeableStateNonVolatile];
+            }
+
+            id<MTLBuffer> buf = [heap->obj newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate offset:heap->offs];
+            if (buf == nil) {
+                GGML_LOG_ERROR("%s: error: failed to create MTLBuffer with size %zu\n", __func__, size_aligned);
+                return nil;
+            }
+
+            heap->n_alloc++;
+            heap->offs += size_aligned;
+
+            [heap->bufs addObject:buf];
+
+            return buf;
+        }
+    }
+
+    // create a new heap that can fit this buffer
+    ggml_metal_heap_ptr * heap_ptr = [ggml_metal_heap_ptr new];
+
+    struct ggml_metal_heap * heap = ggml_metal_heap_init(mem_pool->device, size_aligned);
+    if (heap == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to create heap of size %zu\n", __func__, size_aligned);
+        return NULL;
+    }
+
+    //GGML_LOG_DEBUG("%s: creating new heap of size %zu, got %zu\n", __func__, size_aligned, [heap->obj size]);
+
+    heap_ptr.data = heap;
+    ggml_metal_heap_reset(heap);
+
+    [heap->obj setPurgeableState:MTLPurgeableStateNonVolatile];
+    id<MTLBuffer> buf = [heap->obj newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate offset:heap->offs];
+    if (buf == nil) {
+        GGML_LOG_ERROR("%s: error: failed to create MTLBuffer with size %zu\n", __func__, size_aligned);
+        return NULL;
+    }
+
+    heap->n_alloc++;
+    heap->offs += size_aligned;
+
+    [heap->bufs addObject:buf];
+
+    [mem_pool->heaps addObject:heap_ptr];
+    mem_pool->n_heaps++;
+
+    return buf;
+}
+
+struct ggml_metal_command_buffer {
+    id<MTLCommandBuffer> obj;
+
+    // each command buffer has a memory pool from which it can allocate temporary buffers during the compute
+    struct ggml_metal_mem_pool * mem_pool;
+};
+
 struct ggml_backend_metal_context {
+    id<MTLDevice>       device;
     id<MTLCommandQueue> queue;
 
     dispatch_queue_t d_queue;
@@ -515,7 +767,7 @@ struct ggml_backend_metal_context {
     void (^encode_async)(size_t ith);
 
     // n_cb command buffers + 1 used by the main thread
-    id<MTLCommandBuffer> command_buffers[GGML_METAL_MAX_COMMAND_BUFFERS + 1];
+    struct ggml_metal_command_buffer cmd_bufs[GGML_METAL_MAX_COMMAND_BUFFERS + 1];
 
     // abort ggml_metal_graph_compute if callback returns true
     ggml_abort_callback abort_callback;
@@ -705,9 +957,11 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     struct ggml_backend_metal_device_context * ctx_dev = dev->context;
 
     id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
     GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
-    ctx->queue  = [device newCommandQueue];
+    ctx->device = device;
+    ctx->queue = [device newCommandQueue];
     if (ctx->queue == nil) {
         GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
         return NULL;
@@ -768,7 +1022,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     ctx->gf = nil;
     ctx->encode_async = nil;
     for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
-        ctx->command_buffers[i] = nil;
+        ctx->cmd_bufs[i].obj = nil;
+
+        ctx->cmd_bufs[i].mem_pool = ggml_metal_mem_pool_init();
+        ctx->cmd_bufs[i].mem_pool->device = device;
     }
 
 #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
@@ -1181,6 +1438,12 @@ static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
 
     [ctx->queue release];
 
+    for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
+        // ctx->cmd_bufs[i].obj is auto released
+
+        ggml_metal_mem_pool_free(ctx->cmd_bufs[i].mem_pool);
+    }
+
     dispatch_release(ctx->d_queue);
 
     free(ctx);
@@ -1486,10 +1749,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
     }
 }
 
-static void ggml_metal_encode_node(
+static bool ggml_metal_encode_node(
                         ggml_backend_t   backend,
                                    int   idx,
-          id<MTLComputeCommandEncoder>   encoder) {
+          id<MTLComputeCommandEncoder>   encoder,
+            struct ggml_metal_mem_pool * mem_pool) {
     struct ggml_backend_metal_context        * ctx     = backend->context;
     struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
 
@@ -1505,7 +1769,7 @@ static void ggml_metal_encode_node(
     struct ggml_tensor * dst  = node;
 
     if (ggml_is_empty(dst)) {
-        return;
+        return true;
     }
 
     switch (dst->op) {
@@ -1516,7 +1780,7 @@ static void ggml_metal_encode_node(
         case GGML_OP_PERMUTE:
             {
                 // noop -> next node
-            } return;
+            } return true;
         default:
             {
             } break;
@@ -1527,6 +1791,8 @@ static void ggml_metal_encode_node(
         GGML_ABORT("unsupported op");
     }
 
+    ggml_metal_mem_pool_clear(mem_pool);
+
     const int64_t  ne00 = src0 ? src0->ne[0] : 0;
     const int64_t  ne01 = src0 ? src0->ne[1] : 0;
     const int64_t  ne02 = src0 ? src0->ne[2] : 0;
@@ -2173,26 +2439,76 @@ static void ggml_metal_encode_node(
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                 const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-                ggml_metal_kargs_soft_max args = {
+// use this branch to test the ggml_metal_mem_pool functionality
+#if 0
+                // cpy to tmp buffer in MTLHeap
+
+                id<MTLBuffer> h_src0 = h_src0 = ggml_metal_mem_pool_alloc(mem_pool, ggml_nbytes(src0));
+                if (!h_src0) {
+                    GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, ggml_nbytes(src0));
+                    return false;
+                }
+
+                offs_src0 = 0;
+
+                ggml_metal_kargs_cpy args_cpy = {
                     /*.ne00 =*/ ne00,
                     /*.ne01 =*/ ne01,
                     /*.ne02 =*/ ne02,
-                    /*.scale =*/ scale,
-                    /*.max_bias =*/ max_bias,
-                    /*.m0 =*/ m0,
-                    /*.m1 =*/ m1,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne0  =*/ ne00,
+                    /*.ne1  =*/ ne01,
+                    /*.ne2  =*/ ne02,
+                    /*.ne3  =*/ ne03,
+                    /*.nb0  =*/ nb00,
+                    /*.nb1  =*/ nb01,
+                    /*.nb2  =*/ nb02,
+                    /*.nb3  =*/ nb03,
+                };
+
+                if (src0->type == GGML_TYPE_F16) {
+                    [encoder setComputePipelineState:ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline];
+                } else {
+                    [encoder setComputePipelineState:ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline];
+                }
+                [encoder setBytes:&args_cpy length:sizeof(args_cpy) atIndex:0];
+                [encoder setBuffer:id_src0  offset:offs_src0        atIndex:1];
+                [encoder setBuffer:h_src0   offset:0                atIndex:2];
+
+                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+                int nth_cpy = MIN(1024, ne00 / ggml_blck_size(src0->type));
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth_cpy, 1, 1)];
+
+#else
+                id<MTLBuffer> h_src0 = id_src0;
+#endif
+                // softmax
+
+                ggml_metal_kargs_soft_max args = {
+                    /*.ne00        =*/ ne00,
+                    /*.ne01        =*/ ne01,
+                    /*.ne02        =*/ ne02,
+                    /*.scale       =*/ scale,
+                    /*.max_bias    =*/ max_bias,
+                    /*.m0          =*/ m0,
+                    /*.m1          =*/ m1,
                     /*.n_head_log2 =*/ n_head_log2,
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
+                [encoder setBuffer:h_src0 offset:offs_src0      atIndex:0];
                 if (id_src1) {
-                    [encoder setBuffer:id_src1 offset:offs_src1   atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                 } else {
-                    [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
+                    [encoder setBuffer:h_src0 offset:offs_src0  atIndex:1];
                 }
-                [encoder setBuffer:id_dst      offset:offs_dst            atIndex:2];
-                [encoder setBytes:&args        length:sizeof(args)        atIndex:3];
+                [encoder setBuffer:id_dst offset:offs_dst       atIndex:2];
+                [encoder setBytes:&args   length:sizeof(args)   atIndex:3];
 
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
@@ -4601,6 +4917,8 @@ static void ggml_metal_encode_node(
                 GGML_ABORT("fatal error");
             }
     }
+
+    return true;
 }
 
 static enum ggml_status ggml_metal_graph_compute(
@@ -4654,25 +4972,25 @@ static enum ggml_status ggml_metal_graph_compute(
         }
 
         // the main thread commits the first few commands immediately
-        // command_buffer[n_cb]
+        // cmd_buf[n_cb]
         {
-            id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
-            ctx->command_buffers[n_cb] = command_buffer;
+            id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+            ctx->cmd_bufs[n_cb].obj = cmd_buf;
 
-            [command_buffer enqueue];
+            [cmd_buf enqueue];
             ctx->encode_async(n_cb);
         }
 
         // prepare the rest of the command buffers asynchronously
-        // command_buffer[0.. n_cb)
+        // cmd_buf[0.. n_cb)
         for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-            id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
-            ctx->command_buffers[cb_idx] = command_buffer;
+            id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+            ctx->cmd_bufs[cb_idx].obj = cmd_buf;
 
             // always enqueue the first two command buffers
             // enqueue all of the command buffers if we don't need to abort
             if (cb_idx < 2 || ctx->abort_callback == NULL) {
-                [command_buffer enqueue];
+                [cmd_buf enqueue];
             }
         }
 
@@ -4681,14 +4999,14 @@ static enum ggml_status ggml_metal_graph_compute(
         // wait for completion and check status of each command buffer
         // needed to detect if the device ran out-of-memory for example (#1881)
         {
-            id<MTLCommandBuffer> command_buffer = ctx->command_buffers[n_cb];
-            [command_buffer waitUntilCompleted];
+            id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[n_cb].obj;
+            [cmd_buf waitUntilCompleted];
 
-            MTLCommandBufferStatus status = [command_buffer status];
+            MTLCommandBufferStatus status = [cmd_buf status];
             if (status != MTLCommandBufferStatusCompleted) {
                 GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, n_cb, status);
                 if (status == MTLCommandBufferStatusError) {
-                    GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
+                    GGML_LOG_INFO("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
 
                 return GGML_STATUS_FAILED;
@@ -4696,20 +5014,20 @@ static enum ggml_status ggml_metal_graph_compute(
         }
 
         for (int i = 0; i < n_cb; ++i) {
-            id<MTLCommandBuffer> command_buffer = ctx->command_buffers[i];
-            [command_buffer waitUntilCompleted];
+            id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[i].obj;
+            [cmd_buf waitUntilCompleted];
 
-            MTLCommandBufferStatus status = [command_buffer status];
+            MTLCommandBufferStatus status = [cmd_buf status];
             if (status != MTLCommandBufferStatusCompleted) {
                 GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
                 if (status == MTLCommandBufferStatusError) {
-                    GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
+                    GGML_LOG_INFO("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
 
                 return GGML_STATUS_FAILED;
             }
 
-            id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? ctx->command_buffers[i + 1] : nil);
+            id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? ctx->cmd_bufs[i + 1].obj : nil);
             if (!next_buffer) {
                 continue;
             }
@@ -5092,8 +5410,9 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
 
         const int n_nodes_per_cb = ctx->n_nodes_per_cb;
 
-        id<MTLCommandBuffer> command_buffer  = ctx->command_buffers[cb_idx];
-        id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoder];
+        id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[cb_idx].obj;
+
+        id<MTLComputeCommandEncoder> encoder = [cmd_buf computeCommandEncoder];
 
         int node_start = 0;
         int node_end   = n_nodes_0;
@@ -5105,22 +5424,29 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
 
         const bool should_capture = ctx->capture_next_compute;
 
+        struct ggml_metal_mem_pool * mem_pool = ctx->cmd_bufs[cb_idx].mem_pool;
+        ggml_metal_mem_pool_reset(mem_pool);
+
         for (int idx = node_start; idx < node_end; ++idx) {
             if (should_capture) {
                 [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(ggml_graph_node(ctx->gf, idx)) encoding:NSUTF8StringEncoding]];
             }
 
-            ggml_metal_encode_node(backend, idx, encoder);
+            const bool res = ggml_metal_encode_node(backend, idx, encoder, mem_pool);
 
             if (should_capture) {
                 [encoder popDebugGroup];
             }
+
+            if (!res) {
+                break;
+            }
         }
 
         [encoder endEncoding];
 
         if (cb_idx < 2 || ctx->abort_callback == NULL) {
-            [command_buffer commit];
+            [cmd_buf commit];
         }
     });
 }

ggml/src/ggml-metal/ggml-metal.metal

@@ -3192,7 +3192,7 @@ kernel void kernel_flash_attn_ext(
 
     {
         float S[Q] = { [0 ... Q-1] = 0.0f };
-        float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
+        float M[Q] = { [0 ... Q-1] = -__FLT_MAX__/2 };
 
         // thread indices inside the simdgroup
         // TODO: see if we can utilize quad-group functions for better performance
@@ -3452,7 +3452,7 @@ kernel void kernel_flash_attn_ext(
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
         float S = { 0.0f };
-        float M = { -__FLT16_MAX__/2 };
+        float M = { -__FLT_MAX__/2 };
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
@@ -3699,7 +3699,7 @@ kernel void kernel_flash_attn_ext_vec(
 
     {
         float S = 0.0f;
-        float M = -__FLT16_MAX__/2;
+        float M = -__FLT_MAX__/2;
 
         // thread indices inside the simdgroup
         const short tx = tiisg%NL;

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -378,8 +378,8 @@ static bool parse_endpoint(const std::string & endpoint, std::string & host, int
 }
 
 // RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
-// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
-static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
+// No response
+static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size) {
     uint8_t cmd_byte = cmd;
     if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
         return false;
@@ -390,6 +390,15 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
     if (!send_data(sock->fd, input, input_size)) {
         return false;
     }
+    return true;
+}
+
+// RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
+// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
+static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
+    if (!send_rpc_cmd(sock, cmd, input, input_size)) {
+        return false;
+    }
     // TODO: currently the output_size is always known, do we need support for commands with variable output size?
     // even if we do, we can skip sending output_size from the server for commands with known output size
     uint64_t out_size;
@@ -509,6 +518,11 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
     result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
     result.view_offs = tensor->view_offs;
     result.data = reinterpret_cast<uint64_t>(tensor->data);
+
+    // Avoid sending uninitialized data over the wire
+    memset(result.name, 0, sizeof(result.name));
+    memset(result.padding, 0, sizeof(result.padding));
+
     snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
     return result;
 }
@@ -555,7 +569,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
     memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
     memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size(), nullptr, 0);
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size());
     GGML_ASSERT(status);
 }
 
@@ -973,8 +987,21 @@ bool rpc_server::buffer_clear(const rpc_msg_buffer_clear_req & request) {
 }
 
 ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
+    // Validate tensor type before using it
+    if (tensor->type >= GGML_TYPE_COUNT) {
+        GGML_LOG_ERROR("[%s] invalid tensor type received: %u\n", __func__, tensor->type);
+        return nullptr;
+    }
+
     ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
         tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+
+    // ggml_new_tensor_4d might fail if dimensions are invalid, although less likely to crash than invalid type
+    if (result == nullptr) {
+        GGML_LOG_ERROR("[%s] ggml_new_tensor_4d failed for type %u\\n", __func__, tensor->type);
+        return nullptr;
+    }
+
     for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
         result->nb[i] = tensor->nb[i];
     }
@@ -1034,7 +1061,9 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
         const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
 
         if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
-            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+            GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu) out of buffer bounds [0x%zx, 0x%zx)\n",
+                           __func__, in_tensor->data, offset, size, p0, p1);
+            return false;
         }
     }
 
@@ -1109,7 +1138,9 @@ bool rpc_server::set_tensor_hash(const std::vector<uint8_t> & input, rpc_msg_set
         const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
 
         if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
-            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+            GGML_LOG_ERROR("[%s] tensor data region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%zu, hash=0x%" PRIx64 ") out of buffer bounds [0x%zx, 0x%zx)\n",
+                           __func__, in_tensor->data, offset, size, *hash, p0, p1);
+            return false;
         }
     }
     ggml_backend_tensor_set(tensor, cached_file.data(), offset, size);
@@ -1174,7 +1205,9 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<
         if (request.tensor.data + request.offset < p0 ||
             request.tensor.data + request.offset >= p1 ||
             request.size > (p1 - request.tensor.data - request.offset)) {
-                GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+                GGML_LOG_ERROR("[%s] requested tensor region (data=0x%" PRIx64 ", offset=%" PRIu64 ", size=%" PRIu64 ") out of buffer bounds [0x%zx, 0x%zx)\n",
+                               __func__, request.tensor.data, request.offset, request.size, p0, p1);
+                return false;
         }
     }
 
@@ -1228,22 +1261,50 @@ ggml_tensor * rpc_server::create_node(uint64_t id,
                                       struct ggml_context * ctx,
                                       const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
                                       std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
-    if (id == 0) {
-        return nullptr;
-    }
     if (tensor_map.find(id) != tensor_map.end()) {
         return tensor_map[id];
     }
-    const rpc_tensor * tensor = tensor_ptrs.at(id);
+    // Safely find the tensor pointer
+    auto it_ptr = tensor_ptrs.find(id);
+    if (it_ptr == tensor_ptrs.end()) {
+        return nullptr;
+    }
+    const rpc_tensor * tensor = it_ptr->second;
+
     struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
     if (result == nullptr) {
         return nullptr;
     }
     tensor_map[id] = result;
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+        // Check if the source ID is 0 before calling create_node recursively
+        if (tensor->src[i] == 0) {
+            result->src[i] = nullptr;
+        } else {
+            result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+            // If the recursive call failed for a non-zero ID, propagate the error
+            if (result->src[i] == nullptr) {
+                GGML_LOG_ERROR("[%s] failed to create source node %d (src_id=%" PRIu64 ") for node id %" PRIu64 "\n",
+                               __func__, i, tensor->src[i], id);
+                // Must return nullptr to signal failure up the call stack
+                return nullptr;
+            }
+        }
+    }
+
+    // Handle view_src similarly
+    if (tensor->view_src == 0) {
+        result->view_src = nullptr;
+    } else {
+        result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
+        // If the recursive call failed for a non-zero ID, propagate the error
+        if (result->view_src == nullptr) {
+            GGML_LOG_ERROR("[%s] failed to create view_src node (view_src_id=%" PRIu64 ") for node id %" PRIu64 "\n",
+                           __func__, tensor->view_src, id);
+            // Must return nullptr to signal failure up the call stack
+            return nullptr;
+        }
     }
-    result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
     result->view_offs = tensor->view_offs;
     return result;
 }
@@ -1269,6 +1330,7 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
     GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
 
     size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+
     struct ggml_init_params params = {
         /*.mem_size   =*/ buf_size,
         /*.mem_buffer =*/ NULL,
@@ -1288,6 +1350,14 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, rpc_msg_graph
         int64_t id;
         memcpy(&id, &nodes[i], sizeof(id));
         graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map);
+
+        // Check if create_node failed for a *non-zero* ID.
+        // If id was 0, create_node returning nullptr is expected.
+        // If id was non-zero and create_node returned nullptr, it indicates a deserialization error.
+        if (graph->nodes[i] == nullptr && id != 0) {
+            GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id);
+            return false;
+        }
     }
     ggml_status status = ggml_backend_graph_compute(backend, graph);
     response.result = status;
@@ -1352,7 +1422,9 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                     return;
                 }
                 rpc_msg_get_alloc_size_rsp response;
-                server.get_alloc_size(request, response);
+                if (!server.get_alloc_size(request, response)) {
+                    return;
+                }
                 if (!send_msg(sockfd, &response, sizeof(response))) {
                     return;
                 }
@@ -1428,9 +1500,6 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                 if (!server.set_tensor(input)) {
                     return;
                 }
-                if (!send_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
                 break;
             }
             case RPC_CMD_SET_TENSOR_HASH: {

ggml/src/ggml-sycl/common.hpp

@@ -313,7 +313,6 @@ struct ggml_backend_sycl_context {
     int device;
     std::string name;
     optimize_feature opt_feature;
-    bool optimized_graph=false;
 
     queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } };
 
@@ -494,5 +493,9 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
 
 int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
 
+constexpr size_t ceil_div(const size_t m, const size_t n) {
+    return (m + n - 1) / n;
+}
+
 bool gpu_has_xmx(sycl::device &dev);
 #endif // GGML_SYCL_COMMON_HPP

ggml/src/ggml-sycl/element_wise.cpp

@@ -21,6 +21,27 @@ static void acc_f32(const float * x, const float * y, float * dst, const int ne,
     }
 }
 
+template<typename T>
+static void sgn(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
+    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
+        dst[i] = x[i] > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x[i] < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
+    }
+}
+
+template<typename T>
+static void abs_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
+    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
+        dst[i] = sycl::fabs(x[i]);
+    }
+}
+
+template<typename T>
+static void elu_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
+    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
+        dst[i] = (x[i] > static_cast<T>(0.f)) ? x[i] : sycl::expm1(x[i]);
+    }
+}
+
 template<typename T>
 static void gelu(const T * x, T * dst, const int k,
                      const sycl::nd_item<3> &item_ct1) {
@@ -335,6 +356,37 @@ static void silu_sycl(const T *x, T *dst, const int k,
         });
 }
 
+template<typename T>
+static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
+    // hard code for now
+    const int num_blocks = ceil_div(k, 256);
+    stream->parallel_for(
+            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
+            sgn(x, dst, k, item_ct1);
+            });
+}
+
+template<typename T>
+static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
+    // hard code for now
+    const int num_blocks = ceil_div(k, 256);
+    stream->parallel_for(
+            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
+            abs_op(x, dst, k, item_ct1);
+            });
+}
+
+
+template<typename T>
+static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
+    // hard code for now
+    const int num_blocks = ceil_div(k, 256);
+    stream->parallel_for(
+            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
+            elu_op(x, dst, k, item_ct1);
+            });
+}
+
 template<typename T>
 static void gelu_quick_sycl(const T *x, T *dst, const int k,
                                 queue_ptr stream) {
@@ -574,6 +626,106 @@ static void clamp_sycl(const T *x, T *dst, const float min,
         });
 }
 
+inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+#if defined (GGML_SYCL_F16)
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+
+#else
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+#endif
+    GGML_ASSERT(dst->src[0]->type == dst->type);
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    switch (dst->type) {
+#if defined (GGML_SYCL_F16)
+        case GGML_TYPE_F16:
+            {
+                auto data_pts = cast_data<sycl::half>(dst);
+                sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                break;
+            }
+#endif
+        case GGML_TYPE_F32:
+            {
+                auto data_pts = cast_data<float>(dst);
+                sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                break;
+            }
+        default:
+            GGML_ABORT("GGML tensor type not supported!\n");
+            break;
+    }
+}
+
+inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+#if defined (GGML_SYCL_F16)
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+
+#else
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+#endif
+    GGML_ASSERT(dst->src[0]->type == dst->type);
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    switch (dst->type) {
+#if defined (GGML_SYCL_F16)
+        case GGML_TYPE_F16:
+            {
+                auto data_pts = cast_data<sycl::half>(dst);
+                abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                break;
+            }
+#endif
+        case GGML_TYPE_F32:
+            {
+                auto data_pts = cast_data<float>(dst);
+                abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                break;
+            }
+        default:
+            GGML_ABORT("GGML tensor type not supported!\n");
+            break;
+    }
+}
+
+
+inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+#if defined (GGML_SYCL_F16)
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
+
+#else
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+#endif
+    GGML_ASSERT(dst->src[0]->type == dst->type);
+    dpct::queue_ptr main_stream = ctx.stream();
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    switch (dst->type) {
+#if defined (GGML_SYCL_F16)
+        case GGML_TYPE_F16:
+            {
+                auto data_pts = cast_data<sycl::half>(dst);
+                elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                break;
+            }
+#endif
+        case GGML_TYPE_F32:
+            {
+                auto data_pts = cast_data<float>(dst);
+                elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                break;
+            }
+        default:
+            GGML_ABORT("GGML tensor type not supported!\n");
+            break;
+    }
+}
+
 inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
 #if defined (GGML_SYCL_F16)
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
@@ -1388,3 +1540,20 @@ void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_SYCL_DEBUG("call %s done\n", __func__);
 }
 
+void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type));
+    ggml_sycl_op_sgn(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type));
+    ggml_sycl_op_abs(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type));
+    ggml_sycl_op_elu(ctx, dst);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}

ggml/src/ggml-sycl/element_wise.hpp

@@ -66,5 +66,10 @@ void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
+void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 #endif // GGML_SYCL_ELEMENTWISE_HPP
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -38,6 +38,7 @@
 
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
+#include "ggml-sycl/element_wise.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/sycl_hw.hpp"
@@ -192,7 +193,7 @@ static void ggml_check_sycl() try {
 
     if (!initialized) {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
+        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
         GGML_LOG_INFO("Running with Environment Variables:\n");
@@ -2852,6 +2853,64 @@ static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     }
 }
 
+static void reorder_qw(char *data_device, const int ncols, const int nrows,
+                size_t size, size_t offset, dpct::queue_ptr stream) {
+    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
+    SYCL_CHECK(
+        CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
+            .wait()));
+    GGML_ASSERT((size % sizeof(block_q4_0) == 0));
+    GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
+    int offset_blks = offset / sizeof(block_q4_0);
+    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
+    auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
+
+    stream->parallel_for(
+        size / sizeof(block_q4_0),
+            [=](auto i) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+            const block_q4_0* x = (const block_q4_0*)tmp_buf;
+            const int ib = i;
+
+            for (int j = 0; j < QK4_0/2; j ++)
+            {
+                *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
+            }
+            *(d_ptr + ib) = x[ib].d;
+        });
+
+    sycl::free(tmp_buf, *stream);
+}
+
+static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
+    char*data_device = (char*)src0->data;
+    size_t ncols = src0->ne[0];
+    size_t nrows = src0->ne[1];
+    size_t size = ggml_nbytes(src0);
+
+    reorder_qw(data_device, ncols, nrows, size, 0, stream);
+}
+
+/*
+* This function could be called when the OP (mul_mat) function support reorder optimizition.
+*/
+static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1,
+    ggml_tensor * dst) {
+    if (!g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
+        ctx->opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
+        dst->op == GGML_OP_MUL_MAT &&    //limit to some supported cases of Q4_0, to do for more cases.
+        src0->type == GGML_TYPE_Q4_0 &&
+        src1->ne[2]==1 && src1->ne[3]==1) {
+
+        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
+        if (!extra) return; //only happen in CI/UT permute case.
+
+        if (extra->optimized_feature.reorder) return; //skip the tensor which is handled for reorder.
+
+        reorder_qw(src0, ctx->stream());
+        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
+    }
+}
+
 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 
     const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
@@ -2914,6 +2973,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
+        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
         // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream());
     } else if (use_mul_mat_vec_q) {
@@ -2921,6 +2981,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     } else if (use_mul_mat_q) {
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, true);
     } else {
+        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
     }
 }
@@ -3295,6 +3356,15 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                 case GGML_UNARY_OP_EXP:
                     ggml_sycl_exp(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SGN:
+                    ggml_sycl_sgn(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ABS:
+                    ggml_sycl_abs(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ELU:
+                    ggml_sycl_elu(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -3545,71 +3615,8 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-static void reorder_qw(char *data_device, const int ncols, const int nrows,
-                size_t size, size_t offset, dpct::queue_ptr stream) {
-    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
-    SYCL_CHECK(
-        CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
-            .wait()));
-    GGML_ASSERT((size % sizeof(block_q4_0) == 0));
-    GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
-    int offset_blks = offset / sizeof(block_q4_0);
-    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
-    auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
-
-    stream->parallel_for(
-        size / sizeof(block_q4_0),
-            [=](auto i) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-            const block_q4_0* x = (const block_q4_0*)tmp_buf;
-            const int ib = i;
-
-            for (int j = 0; j < QK4_0/2; j ++)
-            {
-                *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
-            }
-            *(d_ptr + ib) = x[ib].d;
-        });
-
-    sycl::free(tmp_buf, *stream);
-}
-
-static void reorder_qw(ggml_tensor * src0, dpct::queue_ptr stream) {
-    char*data_device = (char*)src0->data;
-    size_t ncols = src0->ne[0];
-    size_t nrows = src0->ne[1];
-    size_t size = ggml_nbytes(src0);
-
-    reorder_qw(data_device, ncols, nrows, size, 0, stream);
-}
-
-static void opt_for_reorder(ggml_tensor * dst, dpct::queue_ptr stream) {
-    ggml_tensor *src0 = dst->src[0];
-    ggml_tensor *src1 = dst->src[1];
-
-    if (dst->op == GGML_OP_MUL_MAT && src0->type == GGML_TYPE_Q4_0 &&
-        src1->ne[2]==1 && src1->ne[3]==1) {
-        reorder_qw(src0, stream);
-        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
-        GGML_ASSERT(extra);
-        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
-    }
-}
-
-static void optimize_graph_once(ggml_cgraph * cgraph, ggml_backend_sycl_context * ctx) {
-    dpct::queue_ptr stream = ctx->stream();
-    if (ctx->optimized_graph) {
-       return;
-    }
-    ctx->optimized_graph = true;
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (ctx->opt_feature.reorder) opt_for_reorder(cgraph->nodes[i], stream);
-    }
-}
-
 static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * sycl_ctx, ggml_cgraph * cgraph) {
     ggml_sycl_set_main_device(sycl_ctx->device);
-    if (!g_ggml_sycl_disable_optimize) optimize_graph_once(cgraph, sycl_ctx);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];
@@ -3840,6 +3847,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_SGN:
+                case GGML_UNARY_OP_ABS:
+                case GGML_UNARY_OP_ELU:
 #if defined (GGML_SYCL_F16)
                     return ggml_is_contiguous(op->src[0]) && (op->type == op->src[0]->type);
 #else

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -71,6 +71,22 @@ if (Vulkan_FOUND)
         add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     endif()
 
+    # Compile a test shader to determine whether GL_EXT_bfloat16 is supported.
+    # If it's not, there will be an error to stderr.
+    # If it's supported, set a define to indicate that we should compile those shaders
+    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
+                    OUTPUT_VARIABLE glslc_output
+                    ERROR_VARIABLE glslc_error)
+
+    if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_bfloat16.*")
+        message(STATUS "GL_EXT_bfloat16 not supported by glslc")
+        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT OFF)
+    else()
+        message(STATUS "GL_EXT_bfloat16 supported by glslc")
+        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT ON)
+        add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    endif()
+
     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
 
@@ -142,6 +158,7 @@ if (Vulkan_FOUND)
                     -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT}
                     -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT}
                     -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT}
+                    -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT=${GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT}
             BUILD_COMMAND ${CMAKE_COMMAND} --build .
             INSTALL_COMMAND ${CMAKE_COMMAND} --install .
             INSTALL_DIR ${CMAKE_BINARY_DIR}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -51,6 +51,24 @@
 
 #include "ggml-vulkan-shaders.hpp"
 
+// remove this once it's more widely available in the SDK
+#if !defined(VK_KHR_shader_bfloat16)
+
+#define VK_KHR_shader_bfloat16 1
+#define VK_KHR_SHADER_BFLOAT16_SPEC_VERSION                          1
+#define VK_KHR_SHADER_BFLOAT16_EXTENSION_NAME                        "VK_KHR_shader_bfloat16"
+#define VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR ((VkStructureType)1000141000)
+#define VK_COMPONENT_TYPE_BFLOAT16_KHR                               ((VkComponentTypeKHR)1000141000)
+
+typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
+    VkStructureType                       sType;
+    void*                                 pNext;
+    VkBool32                              shaderBFloat16Type;
+    VkBool32                              shaderBFloat16DotProduct;
+    VkBool32                              shaderBFloat16CooperativeMatrix;
+} VkPhysicalDeviceShaderBfloat16FeaturesKHR;
+#endif
+
 #define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1))
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
@@ -246,6 +264,7 @@ struct vk_device_struct {
     bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
+    vk::DriverId driver_id;
     vk_device_architecture architecture;
     vk_queue compute_queue;
     vk_queue transfer_queue;
@@ -265,8 +284,9 @@ struct vk_device_struct {
     bool subgroup_require_full_support;
 
     bool coopmat_support;
-    bool coopmat_acc_f32_support;
-    bool coopmat_acc_f16_support;
+    bool coopmat_acc_f32_support {};
+    bool coopmat_acc_f16_support {};
+    bool coopmat_bf16_support {};
     uint32_t coopmat_m;
     uint32_t coopmat_n;
     uint32_t coopmat_k;
@@ -292,6 +312,7 @@ struct vk_device_struct {
 
     vk_matmul_pipeline pipeline_matmul_f32 {};
     vk_matmul_pipeline pipeline_matmul_f32_f16 {};
+    vk_matmul_pipeline pipeline_matmul_bf16 {};
     vk_matmul_pipeline2 pipeline_matmul_f16;
     vk_matmul_pipeline2 pipeline_matmul_f16_f32;
 
@@ -300,6 +321,7 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT];
 
     vk_matmul_pipeline pipeline_matmul_id_f32 {};
+    vk_matmul_pipeline pipeline_matmul_id_bf16 {};
     vk_matmul_pipeline2 pipeline_matmul_id_f16;
     vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
 
@@ -332,8 +354,8 @@ struct vk_device_struct {
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32;
-    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
-    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16;
+    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16, pipeline_cpy_f32_bf16;
+    vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16, pipeline_contig_cpy_f32_bf16;
     vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_norm_f32;
@@ -1740,6 +1762,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         m_warptile_mmq_int = { 128,  64,  64, 32, subgroup_size_8,     32, 2, 2, 2, 1, subgroup_size_8 };
         s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32,       32, 2, 2, 1, 1, subgroup_size_8 };
 
+        // chip specific tuning
+        if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
+            m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
+        }
+
         l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
         m_mmq_wg_denoms = m_wg_denoms = { 64,  64, 1 };
         s_mmq_wg_denoms = s_wg_denoms = { 32,  32, 1 };
@@ -1785,6 +1812,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     if (!device->pipeline_matmul_id_f32) {
         device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     }
+    if (!device->pipeline_matmul_bf16) {
+        device->pipeline_matmul_bf16 = std::make_shared<vk_matmul_pipeline_struct>();
+    }
+    if (!device->pipeline_matmul_id_bf16) {
+        device->pipeline_matmul_id_bf16 = std::make_shared<vk_matmul_pipeline_struct>();
+    }
 
     std::vector<std::future<void>> compiles;
     auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint,
@@ -1894,6 +1927,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT)   \
 
         CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+        }
+#endif
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
@@ -1915,6 +1953,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc,  matmul_iq4_nl_f16,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
 
         CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+        }
+#endif
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
         CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
@@ -1968,6 +2011,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, )
+        }
+#endif
 
         if (device->coopmat_acc_f16_support) {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2016,6 +2064,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+#if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (device->coopmat_bf16_support) {
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+        }
+#endif
 
         if (device->coopmat_acc_f16_support) {
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2098,6 +2151,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2133,6 +2188,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2185,6 +2242,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
@@ -2220,6 +2279,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
 
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+
         CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
@@ -2240,8 +2301,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
         CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
         CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-#undef CREATE_MM
     }
+    // reusing CREATE_MM from the fp32 path
+    if ((device->coopmat2 || device->coopmat_support)
+#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+        && !device->coopmat_bf16_support
+#endif
+        ) {
+        // use scalar tile sizes
+        l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
+        m_warptile = { 128,  64,  64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 };
+        s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 };
+
+        l_wg_denoms = {128, 128, 1 };
+        m_wg_denoms = { 64,  64, 1 };
+        s_wg_denoms = { 32,  32, 1 };
+
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
+    }
+#undef CREATE_MM
 
     // mul mat vec
 
@@ -2260,6 +2339,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f32_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f32_f32_len, mul_mat_vec_bf16_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
@@ -2282,6 +2362,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_BF16][i], "mul_mat_vec_bf16_f16_f32_"+std::to_string(i+1), mul_mat_vec_bf16_f16_f32_len, mul_mat_vec_bf16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
         ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
@@ -2305,6 +2386,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_BF16], "mul_mat_vec_id_bf16_f32", mul_mat_vec_id_bf16_f32_len, mul_mat_vec_id_bf16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
@@ -2350,6 +2432,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     // get_rows
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32",  get_rows_f32_len,  get_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16",  get_rows_f16_len,  get_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_BF16], "get_rows_bf16", get_rows_bf16_len, get_rows_bf16_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -2367,6 +2450,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_BF16], "get_rows_bf16_f32", get_rows_bf16_f32_len, get_rows_bf16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
@@ -2393,7 +2477,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
             ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len,              mul_mat_vec_p021_f16_f32_data,              "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
         }
     }
-    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@@ -2404,10 +2488,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_bf16,"cpy_f32_bf16",cpy_f32_bf16_len,cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_bf16,"contig_cpy_f32_bf16",contig_cpy_f32_bf16_len,contig_cpy_f32_bf16_data,"main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     if (device->float_controls_rte_fp16) {
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1);
         ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1);
@@ -2572,6 +2659,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool coopmat2_support = false;
         device->coopmat_support = false;
         device->integer_dot_product = false;
+        bool bfloat16_support = false;
 
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
@@ -2602,6 +2690,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
                        !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) {
                 device->integer_dot_product = true;
 #endif
+            } else if (strcmp("VK_KHR_shader_bfloat16", properties.extensionName) == 0 &&
+                       !getenv("GGML_VK_DISABLE_BFLOAT16")) {
+                bfloat16_support = true;
             }
         }
 
@@ -2658,6 +2749,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->physical_device.getProperties2(&props2);
         device->properties = props2.properties;
         device->vendor_id = device->properties.vendorID;
+        device->driver_id = driver_props.driverID;
 
         const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
 
@@ -2787,6 +2879,17 @@ static vk_device ggml_vk_get_device(size_t idx) {
         }
 #endif
 
+#if defined(VK_KHR_shader_bfloat16)
+        VkPhysicalDeviceShaderBfloat16FeaturesKHR bfloat16_features {};
+        bfloat16_features.pNext = nullptr;
+        bfloat16_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_BFLOAT16_FEATURES_KHR;
+        if (bfloat16_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&bfloat16_features;
+            last_struct = (VkBaseOutStructure *)&bfloat16_features;
+            device_extensions.push_back("VK_KHR_shader_bfloat16");
+        }
+#endif
+
         VkPhysicalDeviceMaintenance4Features maint4_features {};
         maint4_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_4_FEATURES;
         if (maintenance4_support) {
@@ -2984,6 +3087,25 @@ static vk_device ggml_vk_get_device(size_t idx) {
                     device->coopmat_int_n = prop.NSize;
                     device->coopmat_int_k = prop.KSize;
                 }
+#if defined(VK_KHR_shader_bfloat16) && defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+                if (prop.AType == VK_COMPONENT_TYPE_BFLOAT16_KHR &&
+                    prop.BType == VK_COMPONENT_TYPE_BFLOAT16_KHR &&
+                    prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                    prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                    (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup
+                ) {
+                    // coopmat sizes not set yet
+                    if (device->coopmat_m == 0) {
+                        device->coopmat_bf16_support = true;
+                        device->coopmat_m = prop.MSize;
+                        device->coopmat_n = prop.NSize;
+                        device->coopmat_k = prop.KSize;
+                    } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+                        // Only enable if shape is identical
+                        device->coopmat_bf16_support = true;
+                    }
+                }
+#endif
             }
 
             if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) {
@@ -2991,11 +3113,19 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n");
                 device->coopmat_support = false;
             }
+            if (getenv("GGML_VK_DISABLE_BFLOAT16")) {
+                device->coopmat_bf16_support = false;
+            }
         }
 
         if (device->coopmat_support) {
             device_extensions.push_back("VK_KHR_cooperative_matrix");
         }
+#if defined(VK_KHR_shader_bfloat16)
+        if (device->coopmat_bf16_support) {
+            device_extensions.push_back("VK_KHR_shader_bfloat16");
+        }
+#endif
 #endif
         device->name = GGML_VK_NAME + std::to_string(idx);
 
@@ -3452,6 +3582,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
         return ctx->device->pipeline_matmul_f32_f16;
     }
+    if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) {
+        return ctx->device->pipeline_matmul_bf16;
+    }
     if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_f16_f32.f16acc;
@@ -3523,6 +3656,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -3555,6 +3689,9 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_id_f32;
     }
+    if (src0_type == GGML_TYPE_BF16 && src1_type == GGML_TYPE_BF16) {
+        return ctx->device->pipeline_matmul_id_bf16;
+    }
     if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
         if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
             return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
@@ -3608,6 +3745,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context
     switch (a_type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -4343,6 +4481,13 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_cpy_f16_f16;
         }
     }
+    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_BF16) {
+        if (contig) {
+            return ctx->device->pipeline_contig_cpy_f32_bf16;
+        } else {
+            return ctx->device->pipeline_cpy_f32_bf16;
+        }
+    }
     if (src->type == GGML_TYPE_F32) {
         switch (to) {
         case GGML_TYPE_Q4_0:
@@ -4470,8 +4615,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
     bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0;
@@ -4481,25 +4630,25 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     if (mmp == nullptr) {
         // Fall back to f16 dequant mul mat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
         quantize_y = false;
     }
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
-    const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig);
+    const bool qy_needs_dequant = !quantize_y && ((src1->type != f16_type && !y_f32_kernel) || y_non_contig);
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)));
+    const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)));
     const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
@@ -4520,12 +4669,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     vk_pipeline to_q8_1 = nullptr;
 
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -4942,6 +5091,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t nb01 = src0->nb[1];
     const uint64_t nb02 = src0->nb[2];
 
+    const uint64_t nb12 = src1->nb[2];
+
     // const uint64_t ne10 = src1->ne[0];
     const uint64_t ne11 = src1->ne[1];
     const uint64_t ne12 = src1->ne[2];
@@ -4967,6 +5118,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
+    const uint32_t channel_stride_y = nb12 / sizeof(float);
 
     const uint64_t qx_sz = ggml_nbytes(src0);
     const uint64_t qy_sz = ggml_nbytes(src1);
@@ -4997,7 +5149,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
 
     // compute
-    const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
         { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
@@ -5022,7 +5174,7 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
     // when ne12 and ne13 are one.
     } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) &&
-               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -5090,27 +5242,31 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const bool x_non_contig = (ctx->device->coopmat2 && src0->type == GGML_TYPE_F32) ||
                               !ggml_vk_dim01_contiguous(src0);
     const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              (src0->type == GGML_TYPE_BF16 && src1->type != GGML_TYPE_BF16) ||
                               !ggml_vk_dim01_contiguous(src1);
 
+    // If src0 is BF16, try to use a BF16 x BF16 multiply
+    ggml_type f16_type = src0->type == GGML_TYPE_BF16 ? GGML_TYPE_BF16 : GGML_TYPE_F16;
+
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? f16_type : src1->type, (ggml_prec)dst->op_params[0]);
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
-    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
+    const bool qy_needs_dequant = (src1->type != f16_type && !y_f32_kernel) || y_non_contig;
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
+        mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, f16_type, y_f32_kernel ? GGML_TYPE_F32 : f16_type, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
 
-    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type));
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? f16_type : src0->type));
     const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type);
+    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type);
 
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
@@ -5129,12 +5285,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     vk_pipeline to_fp16_vk_1 = nullptr;
 
     if (x_non_contig) {
-        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, f16_type);
     } else {
         to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
     }
     if (y_non_contig) {
-        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, GGML_TYPE_F16);
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1, nullptr, f16_type);
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
@@ -9220,6 +9376,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (src0_type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9255,10 +9412,15 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (a->ne[3] != b->ne[3]) {
                     return false;
                 }
-                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) ||
+                if (!(ggml_vk_dim01_contiguous(op->src[0]) || op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_BF16) ||
                     !(ggml_vk_dim01_contiguous(op->src[1]) || op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F16)) {
                     return false;
                 }
+                if (op->src[0]->type == GGML_TYPE_BF16 && op->src[1]->type == GGML_TYPE_F16) {
+                    // We currently don't have a bf16 x f16 shader, or an fp16->bf16 copy shader.
+                    // So don't support this combination for now.
+                    return false;
+                }
 
                 return true;
             } break;
@@ -9331,6 +9493,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -9361,6 +9524,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     switch (src1_type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -12,6 +12,9 @@ endif()
 if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
 endif()
+if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+endif()
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)

ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp

@@ -18,7 +18,11 @@ void main() {
     // fast path for when all four iterations are in-bounds
     if (idx + (num_iter-1)*num_threads < p.ne) {
         [[unroll]] for (uint i = 0; i < num_iter; ++i) {
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+
+#if defined(DATA_D_BF16)
+            float f = float(data_a[get_aoffset() + idx]);
+            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
             data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
             data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
@@ -31,7 +35,10 @@ void main() {
                 continue;
             }
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+#if defined(DATA_D_BF16)
+            float f = float(data_a[get_aoffset() + idx]);
+            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
             data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
             data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];

ggml/src/ggml-vulkan/vulkan-shaders/copy.comp

@@ -12,7 +12,10 @@ void main() {
         return;
     }
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
+#if defined(DATA_D_BF16)
+    float f = float(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f));
+#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
     data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
 #else
     data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp

@@ -23,6 +23,12 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_BF16)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1]));
+}
+#endif
+
 #if defined(DATA_A_Q4_0)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
@@ -428,7 +434,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
-#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(0, 0);
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -482,7 +482,7 @@ float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCo
     const uint ib8 = (idx & 0x18) >> 3;  // 0..3
     const uint iqs = 8 * ib32 + ib8;
 
-    const uint8_t qs = bl.block.qs[iqs];
+    const uint qs = bl.block.qs[iqs];
     const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));
 
     const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28));

ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp

@@ -20,9 +20,14 @@ void main() {
     const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
     const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
-#ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
+#if defined(DATA_A_BF16)
+    FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
 #else
-    data_d[d_offset + i00] = data_a[a_offset + i00];
+    FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]);
+#endif
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[d_offset + i00] = D_TYPE(v);
+#else
+    data_d[d_offset + i00] = D_TYPE(v);
 #endif
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -6,7 +6,7 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16) && !defined(DATA_A_BF16)
 #define K_PER_ITER 8
 #else
 #define K_PER_ITER 2

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp

@@ -21,7 +21,9 @@ layout (push_constant) uniform parameter
     uint nrows_x;
     uint row_stride_x;
     uint channel_stride_x;
+    uint channel_stride_y;
     uint channel_x_divisor;
+    uint ne12;
     uint b_offset;
     uint d_offset;
 } p;
@@ -33,6 +35,7 @@ void main() {
     const uint row_x     = gl_GlobalInvocationID.y;
     const uint channel   = gl_GlobalInvocationID.z;
     const uint channel_x = channel / p.channel_x_divisor;
+    const uint channel_y = channel % p.ne12;
 
     const uint nrows_y   = p.ncols_x;
     const uint nrows_dst = p.nrows_x;
@@ -56,7 +59,7 @@ void main() {
                 const uint row_y = col_x;
 
                 const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-                const uint iy = channel*nrows_y + row_y;
+                const uint iy = channel_y*p.channel_stride_y + row_y;
 
                 const vec4 av4 = vec4(data_a_v4[ix / 4]);
                 const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -72,7 +75,7 @@ void main() {
             const uint row_y = col_x;
 
             const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel*nrows_y + row_y;
+            const uint iy = channel_y*p.channel_stride_y + row_y;
 
             const vec4 av4 = vec4(data_a_v4[ix / 4]);
             const vec4 bv4 = vec4(data_b_v4[iy / 4]);
@@ -89,7 +92,7 @@ void main() {
             const uint row_y = col_x;
 
             const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
-            const uint iy = channel*nrows_y + row_y;
+            const uint iy = channel_y*p.channel_stride_y + row_y;
 
             const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -10,6 +10,10 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
 
+#if defined(DATA_A_BF16) && defined(COOPMAT)
+#extension GL_EXT_bfloat16 : enable
+#endif
+
 #ifdef COOPMAT
 #extension GL_KHR_cooperative_matrix : enable
 #extension GL_KHR_memory_scope_semantics : enable
@@ -29,6 +33,10 @@
 #define LOAD_VEC_B 1
 #endif
 
+#if !defined(TO_FLOAT_TYPE)
+#define TO_FLOAT_TYPE FLOAT_TYPE
+#endif
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -202,8 +210,8 @@ void main() {
 #endif
 
 #ifdef COOPMAT
-    coopmat<float16_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
-    coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
+    coopmat<FLOAT_TYPE, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
     coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
 
     [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
@@ -248,6 +256,21 @@ void main() {
                 buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f);
             }
 #endif
+#elif defined(DATA_A_BF16)
+#if LOAD_VEC_A == 4
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = TO_FLOAT_TYPE(data_a[idx].x);
+            buf_a[buf_idx + 1] = TO_FLOAT_TYPE(data_a[idx].y);
+            buf_a[buf_idx + 2] = TO_FLOAT_TYPE(data_a[idx].z);
+            buf_a[buf_idx + 3] = TO_FLOAT_TYPE(data_a[idx].w);
+#else
+            if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
+            } else {
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = TO_FLOAT_TYPE(uint16_t(0));
+            }
+#endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a;
@@ -695,13 +718,13 @@ void main() {
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
 #endif
             const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B;
-            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
-            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
-            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
-            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
+            buf_b[buf_idx + 0] = TO_FLOAT_TYPE(data_b[idx].x);
+            buf_b[buf_idx + 1] = TO_FLOAT_TYPE(data_b[idx].y);
+            buf_b[buf_idx + 2] = TO_FLOAT_TYPE(data_b[idx].z);
+            buf_b[buf_idx + 3] = TO_FLOAT_TYPE(data_b[idx].w);
 #elif !MUL_MAT_ID
             if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
-                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }
@@ -709,7 +732,7 @@ void main() {
             const uint row_i = ic * BN + loadc_b + l;
             if (row_i < _ne1) {
                 const u16vec2 row_idx = row_ids[row_i];
-                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -14,6 +14,9 @@
 #extension GL_EXT_buffer_reference : enable
 #extension GL_KHR_shader_subgroup_ballot : enable
 #extension GL_KHR_shader_subgroup_vote : enable
+#ifdef DATA_A_BF16
+#extension GL_EXT_bfloat16 : enable
+#endif
 
 #include "types.comp"
 
@@ -80,6 +83,12 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #define store_scales(a)
 #endif
 
+#if defined(DATA_A_BF16)
+#define MAT_TYPE bfloat16_t
+#else
+#define MAT_TYPE FLOAT_TYPE
+#endif
+
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
 
@@ -271,8 +280,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose);
@@ -286,8 +295,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover4, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose);
@@ -310,8 +319,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose);
@@ -325,8 +334,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BNover2, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose);
@@ -350,8 +359,8 @@ void main() {
 
                 // Manually partial unroll
                 [[unroll]] for (uint j = 0; j < unroll_count; ++j) {
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                    coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                    coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
                     coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                     coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
@@ -365,8 +374,8 @@ void main() {
                 store_scales(tid);
             }
             while (block_k < end_k) {
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-                coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+                coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
                 coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
                 coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
@@ -405,8 +414,8 @@ void main() {
                 fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false);
             }
 
-            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
-            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+            coopmat<MAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
 
             coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
 #ifdef MUL_MAT_ID

ggml/src/ggml-vulkan/vulkan-shaders/test_bfloat16_support.comp

@@ -0,0 +1,7 @@
+#version 460
+
+#extension GL_EXT_bfloat16 : require
+
+void main()
+{
+}

ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -33,6 +33,19 @@
 #endif
 #endif
 
+#if defined(DATA_A_BF16)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE uint16_t
+#elif LOAD_VEC_A == 4
+#define A_TYPE u16vec4
+#elif LOAD_VEC_A == 8
+#error unsupported
+#endif
+#endif
+
 #define QUANT_K_Q4_0 32
 #define QUANT_R_Q4_0 2
 
@@ -1343,4 +1356,18 @@ void init_iq_shmem(uvec3 wgsize)
 }
 #endif
 
+// returns the bfloat value in the low 16b.
+// See ggml_compute_fp32_to_bf16
+uint32_t fp32_to_bf16(float f)
+{
+    uint32_t u = floatBitsToUint(f);
+    u = (u + (0x7fff + ((u >> 16) & 1))) >> 16;
+    return u;
+}
+
+float bf16_to_fp32(uint32_t u)
+{
+    return uintBitsToFloat(u << 16);
+}
+
 #endif // !defined(GGML_TYPES_COMP)

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -63,7 +63,8 @@ const std::vector<std::string> type_names = {
     "iq3_xxs",
     "iq3_s",
     "iq4_xs",
-    "iq4_nl"
+    "iq4_nl",
+    "bf16",
 };
 
 namespace {
@@ -296,7 +297,6 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
     std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
 
     std::map<std::string, std::string> base_dict = {
-        {"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"},
         {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"},
     };
     std::string shader_name = "matmul";
@@ -318,12 +318,45 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
 
     const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
 
-    // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    auto const &FLOAT_TYPE = [&](const std::string &t) -> std::string {
+        if (t == "bf16") {
+            // scalar path promotes to float
+            if (!coopmat && !coopmat2) {
+                return "float";
+            }
+            return "bfloat16_t";
+        }
+        if (coopmat2 || fp16) {
+            return "float16_t";
+        }
+        return "float";
+    };
 
-    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
-    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+    // Shaders with f16 B_TYPE
+    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+
+    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("f16")}, {"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+
+    // bf16
+    {
+        std::string load_vec_a_unaligned = "1";
+        // For aligned matmul loads
+        std::string load_vec_a = coopmat2 ? "1" : "4";
+
+        // scalar path promotes to float
+        std::string to_float_type = (coopmat || coopmat2) ? "uintBitsToBFloat16EXT" : "bf16_to_fp32";
+
+        // If bfloat16 is not supported, then only compile the scalar (promote to fp32) shader
+#if !defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+        if (!(coopmat || coopmat2))
+#endif
+        {
+            string_to_spv(shader_name + "_bf16_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", "4"}, {"B_TYPE", coopmat2 ? "bfloat16_t" : "u16vec4"},   {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_bf16",         source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE("bf16")}, {"TO_FLOAT_TYPE", to_float_type}, {"DATA_A_BF16", "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                      {"B_TYPE", coopmat2 ? "bfloat16_t" : "uint16_t"},                          {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}),                   fp16, coopmat, coopmat2, f16acc);
+        }
+    }
 
     for (const auto& tname : type_names) {
         std::string load_vec_quant = "2";
@@ -332,26 +365,30 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
         else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl"))
             load_vec_quant = "4";
 
+        if (tname == "bf16") {
+            continue;
+        }
+
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant;
+        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? "1" : load_vec_quant;
         // For aligned matmul loads
-        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant;
+        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16" || tname == "bf16") ? load_vec : load_vec_quant;
 
         // don't generate f32 variants for coopmat2
         if (!coopmat2) {
-            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32",         source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float"},            {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
         if (tname != "f16" && tname != "f32") {
-            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
-            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16",         source_name,  merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
         }
 
 #if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
         if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) {
-            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{"FLOAT_TYPE", FLOAT_TYPE(tname)}, {data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc);
         }
 #endif
     }
@@ -393,6 +430,7 @@ void process_shaders() {
             if (tname == "f32") {
                 continue;
             }
+            if (tname == "bf16") continue;
 
             if (tname == "f16") {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
@@ -417,12 +455,12 @@ void process_shaders() {
         string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
-        if (tname != "f16") {
+        if (tname != "f16" && tname != "bf16") {
             string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
         }
 
         if (!string_ends_with(tname, "_k")) {
-            shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
+            shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";
 
             if (tname == "f16") {
                 string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
@@ -447,9 +485,11 @@ void process_shaders() {
     string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    string_to_spv("cpy_f32_bf16","copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
     string_to_spv("contig_cpy_f32_f32", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("contig_cpy_f32_f16", "contig_copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
     string_to_spv("contig_cpy_f16_f16", "contig_copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    string_to_spv("contig_cpy_f32_bf16","contig_copy.comp",{{"A_TYPE", "float"}, {"D_TYPE", "uint16_t"}, {"DATA_D_BF16", "1"}});
 
     for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
         string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});

ggml/src/ggml.c

@@ -4,6 +4,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-threading.h"
+#include "ggml-cpu.h"
 #include "ggml.h"
 
 // FIXME: required here for quantization functions
@@ -382,58 +383,16 @@ void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
     }
 }
 
-// FIXME: these functions must detect the instruction set at runtime, since they are part of the core ggml library
-//        currently, the ggml_cpu_has_* functions are entirely compile-time
 void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
-    int64_t i = 0;
-#if defined(__F16C__)
-    //if (ggml_cpu_has_f16c()) {
-        for (; i + 7 < n; i += 8) {
-            __m256 x_vec = _mm256_loadu_ps(x + i);
-            __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
-            _mm_storeu_si128((__m128i *)(y + i), y_vec);
-        }
-        for(; i + 3 < n; i += 4) {
-            __m128 x_vec = _mm_loadu_ps(x + i);
-            __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
-            _mm_storel_epi64((__m128i *)(y + i), y_vec);
-        }
-    //}
-#endif
-    for (; i < n; i++) {
+    int i = 0;
+    for (; i < n; ++i) {
         y[i] = GGML_FP32_TO_FP16(x[i]);
     }
 }
 
 void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
-    int64_t i = 0;
-#if defined(__AVX512F__)
-    //if (ggml_cpu_has_avx512()) {
-        for (; i + 16 <= n; i += 16) {
-            _mm512_storeu_ps(y + i,
-                            _mm512_castsi512_ps(
-                                _mm512_slli_epi32(
-                                    _mm512_cvtepu16_epi32(
-                                        _mm256_loadu_si256(
-                                            (const __m256i *)(x + i))),
-                                    16)));
-        }
-    //}
-#endif
-#if defined(__AVX2__)
-    //if (ggml_cpu_has_avx2()) {
-        for (; i + 8 <= n; i += 8) {
-            _mm256_storeu_ps(y + i,
-                            _mm256_castsi256_ps(
-                                _mm256_slli_epi32(
-                                    _mm256_cvtepu16_epi32(
-                                        _mm_loadu_si128(
-                                            (const __m128i *)(x + i))),
-                                    16)));
-        }
-    //}
-#endif
-    for (; i < n; i++) {
+    int i = 0;
+    for (; i < n; ++i) {
         y[i] = GGML_BF16_TO_FP32(x[i]);
     }
 }
@@ -956,6 +915,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CONV_TRANSPOSE_1D",
     "IM2COL",
     "IM2COL_BACK",
+    "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
     "POOL_2D",
@@ -993,7 +953,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1050,6 +1010,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "conv_transpose_1d(x)",
     "im2col(x)",
     "im2col_back(x)",
+    "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
     "pool_2d(x)",
@@ -1087,7 +1048,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1344,6 +1305,13 @@ bool ggml_is_permuted(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1] || tensor->nb[1] > tensor->nb[2] || tensor->nb[2] > tensor->nb[3];
 }
 
+bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor) {
+    return
+        tensor->nb[0] > tensor->nb[2] &&
+        tensor->nb[1] > tensor->nb[0] &&
+        tensor->nb[2] == ggml_type_size(tensor->type);
+}
+
 static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -4050,6 +4018,46 @@ struct ggml_tensor * ggml_conv_2d_dw(
     return result;
 }
 
+// ggml_conv_2d_dw_direct
+
+struct ggml_tensor * ggml_conv_2d_dw_direct(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   stride0,
+        int                   stride1,
+        int                   pad0,
+        int                   pad1,
+        int                   dilation0,
+        int                   dilation1) {
+    GGML_ASSERT(a->ne[2] == 1);
+    GGML_ASSERT(a->ne[3] == b->ne[2]);
+    int64_t ne[4];
+    ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], stride0, pad0, dilation0);
+    ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], stride1, pad1, dilation1);
+    ne[2] = b->ne[2];
+    ne[3] = b->ne[3];
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, b->type, 4, ne);
+
+    if (ggml_is_contiguous_channels(b)) {
+        // Result will be permuted the same way as input (CWHN order)
+        const int64_t type_size = ggml_type_size(result->type);
+        GGML_ASSERT(ggml_blck_size(result->type) == 1);
+        result->nb[0] = result->ne[2] * type_size;
+        result->nb[1] = result->ne[0] * result->nb[0];
+        result->nb[2] = type_size;
+    }
+
+    int32_t params[] = { stride0, stride1, pad0, pad1, dilation0, dilation1 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_CONV_2D_DW;
+    result->src[0] = a;
+    result->src[1] = b;
+    return result;
+}
+
 // ggml_conv_transpose_2d_p0
 
 static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {

gguf-py/gguf/constants.py

@@ -104,6 +104,7 @@ class Keys:
         EXPERT_WEIGHTS_SCALE              = "{arch}.expert_weights_scale"
         EXPERT_WEIGHTS_NORM               = "{arch}.expert_weights_norm"
         EXPERT_GATING_FUNC                = "{arch}.expert_gating_func"
+        MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
         LOGIT_SCALE                       = "{arch}.logit_scale"
         DECODER_START_TOKEN_ID            = "{arch}.decoder_start_token_id"
@@ -230,12 +231,17 @@ class Keys:
         BLOCK_COUNT         = "clip.vision.block_count"
         IMAGE_MEAN          = "clip.vision.image_mean"
         IMAGE_STD           = "clip.vision.image_std"
+        SPATIAL_MERGE_SIZE  = "clip.vision.spatial_merge_size"
         USE_GELU            = "clip.use_gelu"
+        USE_SILU            = "clip.use_silu"
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
             LAYERNORM_EPS   = "clip.vision.attention.layer_norm_epsilon"
 
+        class Projector:
+            SCALE_FACTOR    = "clip.vision.projector.scale_factor"
+
 #
 # recommended mapping of model tensor names for storage in gguf
 #
@@ -263,6 +269,7 @@ class MODEL_ARCH(IntEnum):
     REFACT           = auto()
     BERT             = auto()
     NOMIC_BERT       = auto()
+    NOMIC_BERT_MOE   = auto()
     JINA_BERT_V2     = auto()
     BLOOM            = auto()
     STABLELM         = auto()
@@ -481,9 +488,11 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_OUTPUT         = auto()
     V_ENC_OUTPUT_NORM    = auto()
     V_ENC_FFN_UP         = auto()
+    V_ENC_FFN_GATE       = auto()
     V_ENC_FFN_DOWN       = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
+    V_MM_INP_NORM        = auto()
     V_MM_INP_PROJ        = auto() # gemma3
     V_MM_SOFT_EMB_NORM   = auto() # gemma3
     V_RESMPL_POS_EMBD_K  = auto() # minicpmv
@@ -497,6 +506,8 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_Q_NORM      = auto() # minicpmv
     V_RESMPL_PROJ        = auto() # minicpmv
     V_RESMPL_QUERY       = auto() # minicpmv
+    V_TOK_EMBD_IMG_BREAK = auto() # pixtral
+    V_MM_PATCH_MERGER    = auto() # mistral small 3.1
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -515,6 +526,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.REFACT:           "refact",
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
+    MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.BLOOM:            "bloom",
     MODEL_ARCH.STABLELM:         "stablelm",
@@ -733,10 +745,12 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_ENC_OUTPUT:              "v.blk.{bid}.attn_out",
     MODEL_TENSOR.V_ENC_OUTPUT_NORM:         "v.blk.{bid}.ln2",
     MODEL_TENSOR.V_ENC_FFN_UP:              "v.blk.{bid}.ffn_up",
+    MODEL_TENSOR.V_ENC_FFN_GATE:            "v.blk.{bid}.ffn_gate",
     MODEL_TENSOR.V_ENC_FFN_DOWN:            "v.blk.{bid}.ffn_down",
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
+    MODEL_TENSOR.V_MM_INP_NORM:             "mm.input_norm",
     MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
     MODEL_TENSOR.V_RESMPL_POS_EMBD_K:       "resampler.pos_embd_k",
     MODEL_TENSOR.V_RESMPL_ATTN_Q:           "resampler.attn.q",
@@ -749,6 +763,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_RESMPL_Q_NORM:           "resampler.ln_q",
     MODEL_TENSOR.V_RESMPL_PROJ:             "resampler.proj",
     MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
+    MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
+    MODEL_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -767,10 +783,12 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_ENC_OUTPUT,
         MODEL_TENSOR.V_ENC_OUTPUT_NORM,
         MODEL_TENSOR.V_ENC_FFN_UP,
+        MODEL_TENSOR.V_ENC_FFN_GATE,
         MODEL_TENSOR.V_ENC_FFN_DOWN,
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
+        MODEL_TENSOR.V_MM_INP_NORM,
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
         MODEL_TENSOR.V_RESMPL_POS_EMBD_K,
         MODEL_TENSOR.V_RESMPL_ATTN_Q,
@@ -783,6 +801,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_RESMPL_Q_NORM,
         MODEL_TENSOR.V_RESMPL_PROJ,
         MODEL_TENSOR.V_RESMPL_QUERY,
+        MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
+        MODEL_TENSOR.V_MM_PATCH_MERGER,
     ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -950,6 +970,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NOMIC_BERT_MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -2122,6 +2158,12 @@ class GGUFValueType(IntEnum):
             raise ValueError(f"Unknown type: {type(val)}")
 
 
+class VisionProjectorType:
+    GEMMA3 = "gemma3"
+    IDEFICS3 = "idefics3"
+    PIXTRAL = "pixtral"
+
+
 # Items here are (block size, type size)
 QK_K = 256
 GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {

gguf-py/gguf/gguf_writer.py

@@ -728,6 +728,9 @@ class GGUFWriter:
     def add_expert_gating_func(self, value: ExpertGatingFuncType) -> None:
         self.add_uint32(Keys.LLM.EXPERT_GATING_FUNC.format(arch=self.arch), value.value)
 
+    def add_moe_every_n_layers(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value)
+
     def add_swin_norm(self, value: bool) -> None:
         self.add_bool(Keys.LLM.SWIN_NORM.format(arch=self.arch), value)
 
@@ -931,6 +934,56 @@ class GGUFWriter:
     def add_eom_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOM_ID, id)
 
+    # for vision models
+
+    def add_vision_projection_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value)
+
+    def add_vision_has_vision_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.ClipVision.HAS_VISION_ENCODER, value)
+
+    def add_vision_patch_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.PATCH_SIZE, value)
+
+    def add_vision_embedding_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.EMBEDDING_LENGTH, value)
+
+    def add_vision_feed_forward_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.FEED_FORWARD_LENGTH, value)
+
+    def add_vision_block_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.BLOCK_COUNT, value)
+
+    def add_vision_head_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT, value)
+
+    def add_vision_projector_type(self, value: str) -> None:
+        self.add_string(Keys.ClipVision.PROJECTOR_TYPE, value)
+
+    def add_vision_attention_layernorm_eps(self, value: float) -> None:
+        self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)
+
+    def add_vision_image_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.IMAGE_SIZE, value)
+
+    def add_vision_image_mean(self, values: Sequence[float]) -> None:
+        self.add_array(Keys.ClipVision.IMAGE_MEAN, values)
+
+    def add_vision_image_std(self, values: Sequence[float]) -> None:
+        self.add_array(Keys.ClipVision.IMAGE_STD, values)
+
+    def add_vision_spatial_merge_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.SPATIAL_MERGE_SIZE, value)
+
+    def add_vision_use_gelu(self, value: bool) -> None:
+        self.add_bool(Keys.ClipVision.USE_GELU, value)
+
+    def add_vision_use_silu(self, value: bool) -> None:
+        self.add_bool(Keys.ClipVision.USE_SILU, value)
+
+    def add_vision_projector_scale_factor(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/tensor_mapping.py

@@ -290,6 +290,7 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
             "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
             "language_model.model.layers.{bid}.feed_forward.router", # llama4
+            "encoder.layers.{bid}.mlp.router.layer",            # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -322,6 +323,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
+            "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
             "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
@@ -337,6 +339,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.experts.up_proj",         # qwen2moe olmoe (merged)
             "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
             "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w1",        # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
@@ -418,6 +421,7 @@ class TensorNameMap:
             "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
             "model.layers.{bid}.block_sparse_moe.experts.w2",    # phimoe (merged)
             "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4
+            "encoder.layers.{bid}.mlp.experts.mlp.w2",           # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (
@@ -914,6 +918,7 @@ class TensorNameMap:
             "vision_tower.vision_model.embeddings.patch_embedding",
             "vpm.embeddings.patch_embedding",
             "model.vision_model.embeddings.patch_embedding", # SmolVLM
+            "vision_tower.patch_conv", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
@@ -926,52 +931,65 @@ class TensorNameMap:
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj",
             "vpm.encoder.layers.{bid}.self_attn.q_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_K: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_V: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
             "vpm.encoder.layers.{bid}.self_attn.v_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm1",
             "vpm.encoder.layers.{bid}.layer_norm1",
             "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj",
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2",
             "vpm.encoder.layers.{bid}.layer_norm2",
             "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
             "vpm.encoder.layers.{bid}.mlp.fc1",
-            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM
+            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
+            "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
+        ),
+
+        MODEL_TENSOR.V_ENC_FFN_GATE: (
+            "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_FFN_DOWN: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2",
             "vpm.encoder.layers.{bid}.mlp.fc2",
-            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM
+            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
+            "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_PRE_NORM: (
             "vision_tower.vision_model.pre_layrnorm",
+            "vision_tower.ln_pre", # pixtral
         ),
 
         MODEL_TENSOR.V_POST_NORM: (
@@ -983,6 +1001,10 @@ class TensorNameMap:
             "multi_modal_projector.mm_input_projection",
         ),
 
+        MODEL_TENSOR.V_MM_INP_NORM: (
+            "multi_modal_projector.norm",
+        ),
+
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
             "multi_modal_projector.mm_soft_emb_norm",
         ),
@@ -1030,6 +1052,14 @@ class TensorNameMap:
         MODEL_TENSOR.V_RESMPL_QUERY: (
             "resampler.query",
         ),
+
+        MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: (
+            "v.token_embd.img_break", # for pixtral, this is a generated vector
+        ),
+
+        MODEL_TENSOR.V_MM_PATCH_MERGER: (
+            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
+        ),
     }
 
     # architecture-specific block mappings

grammars/README.md

@@ -112,7 +112,7 @@ You can use GBNF grammars:
 
 - In [llama-server](../examples/server)'s completion endpoints, passed as the `grammar` body field
 - In [llama-cli](../examples/main), passed as the `--grammar` & `--grammar-file` flags
-- With [llama-gbnf-validator](../examples/gbnf-validator) tool, to test them against strings.
+- With [test-gbnf-validator](../tests/test-gbnf-validator.cpp), to test them against strings.
 
 ## JSON Schemas → GBNF
 

include/llama.h

@@ -111,6 +111,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
+        LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
     };
 
     enum llama_rope_type {
@@ -1231,6 +1232,7 @@ extern "C" {
         "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)");
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    /// Setting k <= 0 makes this a noop
     LLAMA_API struct llama_sampler * llama_sampler_init_top_k      (int32_t k);
 
     /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751